Merge branch 'main' of github.com:crewAIInc/crewAI into lorenze/feat/plan-execute-pattern

* fix: adjust executor listener value to avoid recursion

* fix: clear call count to ensure zero state

* feat: expose max method call kwarg

.env.test

@@ -21,7 +21,6 @@ OPENROUTER_API_KEY=fake-openrouter-key
 AWS_ACCESS_KEY_ID=fake-aws-access-key
 AWS_SECRET_ACCESS_KEY=fake-aws-secret-key
 AWS_DEFAULT_REGION=us-east-1
-AWS_REGION_NAME=us-east-1
 
 # -----------------------------------------------------------------------------
 # Azure OpenAI Configuration

.github/workflows/publish.yml

@@ -1,8 +1,6 @@
 name: Publish to PyPI
 
 on:
-  repository_dispatch:
-    types: [deployment-tests-passed]
   workflow_dispatch:
     inputs:
       release_tag:
@@ -20,11 +18,8 @@ jobs:
       - name: Determine release tag
         id: release
         run: |
-          # Priority: workflow_dispatch input > repository_dispatch payload > default branch
           if [ -n "${{ inputs.release_tag }}" ]; then
             echo "tag=${{ inputs.release_tag }}" >> $GITHUB_OUTPUT
-          elif [ -n "${{ github.event.client_payload.release_tag }}" ]; then
-            echo "tag=${{ github.event.client_payload.release_tag }}" >> $GITHUB_OUTPUT
           else
             echo "tag=" >> $GITHUB_OUTPUT
           fi

.github/workflows/trigger-deployment-tests.yml

@@ -1,18 +0,0 @@
-name: Trigger Deployment Tests
-
-on:
-  release:
-    types: [published]
-
-jobs:
-  trigger:
-    name: Trigger deployment tests
-    runs-on: ubuntu-latest
-    steps:
-      - name: Trigger deployment tests
-        uses: peter-evans/repository-dispatch@v3
-        with:
-          token: ${{ secrets.CREWAI_DEPLOYMENTS_PAT }}
-          repository: ${{ secrets.CREWAI_DEPLOYMENTS_REPOSITORY }}
-          event-type: crewai-release
-          client-payload: '{"release_tag": "${{ github.event.release.tag_name }}", "release_name": "${{ github.event.release.name }}"}'

conftest.py

@@ -12,6 +12,7 @@ from dotenv import load_dotenv
 import pytest
 from vcr.request import Request  # type: ignore[import-untyped]
 
+
 try:
     import vcr.stubs.httpx_stubs as httpx_stubs  # type: ignore[import-untyped]
 except ModuleNotFoundError:

docs/en/changelog.mdx

@@ -4,6 +4,138 @@ description: "Product updates, improvements, and bug fixes for CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="Mar 04, 2026">
+  ## v1.10.1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## What's Changed
+
+  ### Features
+  - Upgrade Gemini GenAI
+
+  ### Bug Fixes
+  - Adjust executor listener value to avoid recursion
+  - Group parallel function response parts in a single Content object in Gemini
+  - Surface thought output from thinking models in Gemini
+  - Load MCP and platform tools when agent tools are None
+  - Support Jupyter environments with running event loops in A2A
+  - Use anonymous ID for ephemeral traces
+  - Conditionally pass plus header
+  - Skip signal handler registration in non-main threads for telemetry
+  - Inject tool errors as observations and resolve name collisions
+  - Upgrade pypdf from 4.x to 6.7.4 to resolve Dependabot alerts
+  - Resolve critical and high Dependabot security alerts
+
+  ### Documentation
+  - Sync Composio tool documentation across locales
+
+  ## Contributors
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
+<Update label="Feb 27, 2026">
+  ## v1.10.1a1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## What's Changed
+
+  ### Features
+  - Implement asynchronous invocation support in step callback methods
+  - Implement lazy loading for heavy dependencies in Memory module
+
+  ### Documentation
+  - Update changelog and version for v1.10.0
+
+  ### Refactoring
+  - Refactor step callback methods to support asynchronous invocation
+  - Refactor to implement lazy loading for heavy dependencies in Memory module
+
+  ### Bug Fixes
+  - Fix branch for release notes
+
+  ## Contributors
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="Feb 27, 2026">
+  ## v1.10.1a1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## What's Changed
+
+  ### Refactoring
+  - Refactor step callback methods to support asynchronous invocation
+  - Implement lazy loading for heavy dependencies in Memory module
+
+  ### Documentation
+  - Update changelog and version for v1.10.0
+
+  ### Bug Fixes
+  - Make branch for release notes
+
+  ## Contributors
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="Feb 26, 2026">
+  ## v1.10.0
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.0)
+
+  ## What's Changed
+
+  ### Features
+  - Enhance MCP tool resolution and related events
+  - Update lancedb version and add lance-namespace packages
+  - Enhance JSON argument parsing and validation in CrewAgentExecutor and BaseTool
+  - Migrate CLI HTTP client from requests to httpx
+  - Add versioned documentation
+  - Add yanked detection for version notes
+  - Implement user input handling in Flows
+  - Enhance HITL self-loop functionality in human feedback integration tests
+  - Add started_event_id and set in eventbus
+  - Auto update tools.specs
+
+  ### Bug Fixes
+  - Validate tool kwargs even when empty to prevent cryptic TypeError
+  - Preserve null types in tool parameter schemas for LLM
+  - Map output_pydantic/output_json to native structured output
+  - Ensure callbacks are ran/awaited if promise
+  - Capture method name in exception context
+  - Preserve enum type in router result; improve types
+  - Fix cyclic flows silently breaking when persistence ID is passed in inputs
+  - Correct CLI flag format from --skip-provider to --skip_provider
+  - Ensure OpenAI tool call stream is finalized
+  - Resolve complex schema $ref pointers in MCP tools
+  - Enforce additionalProperties=false in schemas
+  - Reject reserved script names for crew folders
+  - Resolve race condition in guardrail event emission test
+
+  ### Documentation
+  - Add litellm dependency note for non-native LLM providers
+  - Clarify NL2SQL security model and hardening guidance
+  - Add 96 missing actions across 9 integrations
+
+  ### Refactoring
+  - Refactor crew to provider
+  - Extract HITL to provider pattern
+  - Improve hook typing and registration
+
+  ## Contributors
+
+  @dependabot[bot], @github-actions[bot], @github-code-quality[bot], @greysonlalonde, @heitorado, @hobostay, @joaomdmoura, @johnvan7, @jonathansampson, @lorenzejay, @lucasgomide, @mattatcha, @mplachta, @nicoferdi96, @theCyberTech, @thiagomoretto, @vinibrsl
+
+</Update>
+
 <Update label="Jan 26, 2026">
   ## v1.9.0
 

docs/en/concepts/llms.mdx

@@ -106,6 +106,15 @@ There are different places in CrewAI code where you can specify the model to use
   </Tab>
 </Tabs>
 
+<Info>
+  CrewAI provides native SDK integrations for OpenAI, Anthropic, Google (Gemini API), Azure, and AWS Bedrock — no extra install needed beyond the provider-specific extras (e.g. `uv add "crewai[openai]"`).
+
+  All other providers are powered by **LiteLLM**. If you plan to use any of them, add it as a dependency to your project:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+</Info>
+
 ## Provider Configuration Examples
 
 CrewAI supports a multitude of LLM providers, each offering unique features, authentication methods, and model capabilities.
@@ -275,6 +284,11 @@ In this section, you'll find detailed examples that help you select, configure,
     | `meta_llama/Llama-4-Maverick-17B-128E-Instruct-FP8` | 128k | 4028 | Text, Image | Text |
     | `meta_llama/Llama-3.3-70B-Instruct` | 128k | 4028 | Text | Text |
     | `meta_llama/Llama-3.3-8B-Instruct` | 128k | 4028 | Text | Text |
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Anthropic">
@@ -470,7 +484,7 @@ In this section, you'll find detailed examples that help you select, configure,
       To get an Express mode API key:
       - New Google Cloud users: Get an [express mode API key](https://cloud.google.com/vertex-ai/generative-ai/docs/start/quickstart?usertype=apikey)
       - Existing Google Cloud users: Get a [Google Cloud API key bound to a service account](https://cloud.google.com/docs/authentication/api-keys)
-      
+
       For more details, see the [Vertex AI Express mode documentation](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/start/quickstart?usertype=apikey).
     </Info>
 
@@ -571,6 +585,11 @@ In this section, you'll find detailed examples that help you select, configure,
     | gemini-1.5-flash               | 1M tokens      | Balanced multimodal model, good for most tasks                    |
     | gemini-1.5-flash-8B            | 1M tokens      | Fastest, most cost-efficient, good for high-frequency tasks       |
     | gemini-1.5-pro                 | 2M tokens      | Best performing, wide variety of reasoning tasks including logical reasoning, coding, and creative collaboration |
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Azure">
@@ -652,6 +671,7 @@ In this section, you'll find detailed examples that help you select, configure,
     # Optional
     AWS_SESSION_TOKEN=<your-session-token>  # For temporary credentials
     AWS_DEFAULT_REGION=<your-region>  # Defaults to us-east-1
+    AWS_REGION_NAME=<your-region>  # Alternative configuration for backwards compatibility with LiteLLM. Defaults to us-east-1
     ```
 
     **Basic Usage:**
@@ -695,6 +715,7 @@ In this section, you'll find detailed examples that help you select, configure,
     - `AWS_SECRET_ACCESS_KEY`: AWS secret key (required)
     - `AWS_SESSION_TOKEN`: AWS session token for temporary credentials (optional)
     - `AWS_DEFAULT_REGION`: AWS region (defaults to `us-east-1`)
+    - `AWS_REGION_NAME`: AWS region (defaults to `us-east-1`). Alternative configuration for backwards compatibility with LiteLLM
 
     **Features:**
     - Native tool calling support via Converse API
@@ -764,6 +785,11 @@ In this section, you'll find detailed examples that help you select, configure,
         model="sagemaker/<my-endpoint>"
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Mistral">
@@ -779,6 +805,11 @@ In this section, you'll find detailed examples that help you select, configure,
         temperature=0.7
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Nvidia NIM">
@@ -865,6 +896,11 @@ In this section, you'll find detailed examples that help you select, configure,
     | rakuten/rakutenai-7b-instruct                                     | 1,024 tokens   | Advanced state-of-the-art LLM with language understanding, superior reasoning, and text generation. |
     | rakuten/rakutenai-7b-chat                                         | 1,024 tokens   | Advanced state-of-the-art LLM with language understanding, superior reasoning, and text generation. |
     | baichuan-inc/baichuan2-13b-chat                                  | 4,096 tokens   | Support Chinese and English chat, coding, math, instruction following, solving quizzes |
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Local NVIDIA NIM Deployed using WSL2">
@@ -905,6 +941,11 @@ In this section, you'll find detailed examples that help you select, configure,
 
         # ...
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Groq">
@@ -926,6 +967,11 @@ In this section, you'll find detailed examples that help you select, configure,
     | Llama 3.1 70B/8B  | 131,072 tokens   | High-performance, large context tasks      |
     | Llama 3.2 Series  | 8,192 tokens     | General-purpose tasks                      |
     | Mixtral 8x7B      | 32,768 tokens    | Balanced performance and context           |
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="IBM watsonx.ai">
@@ -948,6 +994,11 @@ In this section, you'll find detailed examples that help you select, configure,
         base_url="https://api.watsonx.ai/v1"
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Ollama (Local LLMs)">
@@ -961,6 +1012,11 @@ In this section, you'll find detailed examples that help you select, configure,
         base_url="http://localhost:11434"
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Fireworks AI">
@@ -976,6 +1032,11 @@ In this section, you'll find detailed examples that help you select, configure,
         temperature=0.7
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Perplexity AI">
@@ -991,6 +1052,11 @@ In this section, you'll find detailed examples that help you select, configure,
         base_url="https://api.perplexity.ai/"
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Hugging Face">
@@ -1005,6 +1071,11 @@ In this section, you'll find detailed examples that help you select, configure,
         model="huggingface/meta-llama/Meta-Llama-3.1-8B-Instruct"
     )
     ```
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="SambaNova">
@@ -1028,6 +1099,11 @@ In this section, you'll find detailed examples that help you select, configure,
     | Llama 3.2 Series   | 8,192 tokens           | General-purpose, multimodal tasks            |
     | Llama 3.3 70B      | Up to 131,072 tokens   | High-performance and output quality          |
     | Qwen2 familly      | 8,192 tokens           | High-performance and output quality          |
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Cerebras">
@@ -1053,6 +1129,11 @@ In this section, you'll find detailed examples that help you select, configure,
       - Good balance of speed and quality
       - Support for long context windows
     </Info>
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Open Router">
@@ -1075,6 +1156,11 @@ In this section, you'll find detailed examples that help you select, configure,
       - openrouter/deepseek/deepseek-r1
       - openrouter/deepseek/deepseek-chat
     </Info>
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Nebius AI Studio">
@@ -1097,6 +1183,11 @@ In this section, you'll find detailed examples that help you select, configure,
       - Competitive pricing
       - Good balance of speed and quality
     </Info>
+
+    **Note:** This provider uses LiteLLM. Add it as a dependency to your project:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 </AccordionGroup>
 

docs/en/enterprise/features/flow-hitl-management.mdx

@@ -38,22 +38,21 @@ CrewAI Enterprise provides a comprehensive Human-in-the-Loop (HITL) management s
 Configure human review checkpoints within your Flows using the `@human_feedback` decorator. When execution reaches a review point, the system pauses, notifies the assignee via email, and waits for a response.
 
 ```python
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 
 class ContentApprovalFlow(Flow):
     @start()
     def generate_content(self):
-        # AI generates content
         return "Generated marketing copy for Q1 campaign..."
 
-    @listen(generate_content)
     @human_feedback(
         message="Please review this content for brand compliance:",
         emit=["approved", "rejected", "needs_revision"],
     )
-    def review_content(self, content):
-        return content
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "Marketing copy for review..."
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
@@ -62,10 +61,6 @@ class ContentApprovalFlow(Flow):
     @listen("rejected")
     def archive_content(self, result: HumanFeedbackResult):
         print(f"Content rejected. Reason: {result.feedback}")
-
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        print(f"Revision requested: {result.feedback}")
 ```
 
 For complete implementation details, see the [Human Feedback in Flows](/en/learn/human-feedback-in-flows) guide.

docs/en/enterprise/guides/deploy-to-amp.mdx

@@ -177,6 +177,11 @@ You need to push your crew to a GitHub repository. If you haven't created a crew
       ![Set Environment Variables](/images/enterprise/set-env-variables.png)
     </Frame>
 
+    <Info>
+      Using private Python packages? You'll need to add your registry credentials here too.
+      See [Private Package Registries](/en/enterprise/guides/private-package-registry) for the required variables.
+    </Info>
+
   </Step>
 
   <Step title="Deploy Your Crew">

docs/en/enterprise/guides/prepare-for-deployment.mdx

@@ -256,6 +256,12 @@ Before deployment, ensure you have:
 1. **LLM API keys** ready (OpenAI, Anthropic, Google, etc.)
 2. **Tool API keys** if using external tools (Serper, etc.)
 
+<Info>
+  If your project depends on packages from a **private PyPI registry**, you'll also need to configure
+  registry authentication credentials as environment variables. See the
+  [Private Package Registries](/en/enterprise/guides/private-package-registry) guide for details.
+</Info>
+
 <Tip>
   Test your project locally with the same environment variables before deploying
   to catch configuration issues early.

docs/en/enterprise/guides/private-package-registry.mdx

@@ -0,0 +1,263 @@
+---
+title: "Private Package Registries"
+description: "Install private Python packages from authenticated PyPI registries in CrewAI AMP"
+icon: "lock"
+mode: "wide"
+---
+
+<Note>
+  This guide covers how to configure your CrewAI project to install Python packages
+  from private PyPI registries (Azure DevOps Artifacts, GitHub Packages, GitLab, AWS CodeArtifact, etc.)
+  when deploying to CrewAI AMP.
+</Note>
+
+## When You Need This
+
+If your project depends on internal or proprietary Python packages hosted on a private registry
+rather than the public PyPI, you'll need to:
+
+1. Tell UV **where** to find the package (an index URL)
+2. Tell UV **which** packages come from that index (a source mapping)
+3. Provide **credentials** so UV can authenticate during install
+
+CrewAI AMP uses [UV](https://docs.astral.sh/uv/) for dependency resolution and installation.
+UV supports authenticated private registries through `pyproject.toml` configuration combined
+with environment variables for credentials.
+
+## Step 1: Configure pyproject.toml
+
+Three pieces work together in your `pyproject.toml`:
+
+### 1a. Declare the dependency
+
+Add the private package to your `[project.dependencies]` like any other dependency:
+
+```toml
+[project]
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+```
+
+### 1b. Define the index
+
+Register your private registry as a named index under `[[tool.uv.index]]`:
+
+```toml
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+```
+
+<Info>
+  The `name` field is important — UV uses it to construct the environment variable names
+  for authentication (see [Step 2](#step-2-set-authentication-credentials) below).
+
+  Setting `explicit = true` means UV won't search this index for every package — only the
+  ones you explicitly map to it in `[tool.uv.sources]`. This avoids unnecessary queries
+  against your private registry and protects against dependency confusion attacks.
+</Info>
+
+### 1c. Map the package to the index
+
+Tell UV which packages should be resolved from your private index using `[tool.uv.sources]`:
+
+```toml
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+### Complete example
+
+```toml
+[project]
+name = "my-crew-project"
+version = "0.1.0"
+requires-python = ">=3.10,<=3.13"
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+
+[tool.crewai]
+type = "crew"
+
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+After updating `pyproject.toml`, regenerate your lock file:
+
+```bash
+uv lock
+```
+
+<Warning>
+  Always commit the updated `uv.lock` along with your `pyproject.toml` changes.
+  The lock file is required for deployment — see [Prepare for Deployment](/en/enterprise/guides/prepare-for-deployment).
+</Warning>
+
+## Step 2: Set Authentication Credentials
+
+UV authenticates against private indexes using environment variables that follow a naming convention
+based on the index name you defined in `pyproject.toml`:
+
+```
+UV_INDEX_{UPPER_NAME}_USERNAME
+UV_INDEX_{UPPER_NAME}_PASSWORD
+```
+
+Where `{UPPER_NAME}` is your index name converted to **uppercase** with **hyphens replaced by underscores**.
+
+For example, an index named `my-private-registry` uses:
+
+| Variable | Value |
+|----------|-------|
+| `UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME` | Your registry username or token name |
+| `UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD` | Your registry password or token/PAT |
+
+<Warning>
+  These environment variables **must** be added via the CrewAI AMP **Environment Variables** settings —
+  either globally or at the deployment level. They cannot be set in `.env` files or hardcoded in your project.
+
+  See [Setting Environment Variables in AMP](#setting-environment-variables-in-amp) below.
+</Warning>
+
+## Registry Provider Reference
+
+The table below shows the index URL format and credential values for common registry providers.
+Replace placeholder values with your actual organization and feed details.
+
+| Provider | Index URL | Username | Password |
+|----------|-----------|----------|----------|
+| **Azure DevOps Artifacts** | `https://pkgs.dev.azure.com/{org}/_packaging/{feed}/pypi/simple/` | Any non-empty string (e.g. `token`) | Personal Access Token (PAT) with Packaging Read scope |
+| **GitHub Packages** | `https://pypi.pkg.github.com/{owner}/simple/` | GitHub username | Personal Access Token (classic) with `read:packages` scope |
+| **GitLab Package Registry** | `https://gitlab.com/api/v4/projects/{project_id}/packages/pypi/simple/` | `__token__` | Project or Personal Access Token with `read_api` scope |
+| **AWS CodeArtifact** | Use the URL from `aws codeartifact get-repository-endpoint` | `aws` | Token from `aws codeartifact get-authorization-token` |
+| **Google Artifact Registry** | `https://{region}-python.pkg.dev/{project}/{repo}/simple/` | `_json_key_base64` | Base64-encoded service account key |
+| **JFrog Artifactory** | `https://{instance}.jfrog.io/artifactory/api/pypi/{repo}/simple/` | Username or email | API key or identity token |
+| **Self-hosted (devpi, Nexus, etc.)** | Your registry's simple API URL | Registry username | Registry password |
+
+<Tip>
+  For **AWS CodeArtifact**, the authorization token expires periodically.
+  You'll need to refresh the `UV_INDEX_*_PASSWORD` value when it expires.
+  Consider automating this in your CI/CD pipeline.
+</Tip>
+
+## Setting Environment Variables in AMP
+
+Private registry credentials must be configured as environment variables in CrewAI AMP.
+You have two options:
+
+<Tabs>
+  <Tab title="Web Interface">
+    1. Log in to [CrewAI AMP](https://app.crewai.com)
+    2. Navigate to your automation
+    3. Open the **Environment Variables** tab
+    4. Add each variable (`UV_INDEX_*_USERNAME` and `UV_INDEX_*_PASSWORD`) with its value
+
+    See the [Deploy to AMP — Set Environment Variables](/en/enterprise/guides/deploy-to-amp#set-environment-variables) step for details.
+  </Tab>
+  <Tab title="CLI Deployment">
+    Add the variables to your local `.env` file before running `crewai deploy create`.
+    The CLI will securely transfer them to the platform:
+
+    ```bash
+    # .env
+    OPENAI_API_KEY=sk-...
+    UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+    UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat-here
+    ```
+
+    ```bash
+    crewai deploy create
+    ```
+  </Tab>
+</Tabs>
+
+<Warning>
+  **Never** commit credentials to your repository. Use AMP environment variables for all secrets.
+  The `.env` file should be listed in `.gitignore`.
+</Warning>
+
+To update credentials on an existing deployment, see [Update Your Crew — Environment Variables](/en/enterprise/guides/update-crew).
+
+## How It All Fits Together
+
+When CrewAI AMP builds your automation, the resolution flow works like this:
+
+<Steps>
+  <Step title="Build starts">
+    AMP pulls your repository and reads `pyproject.toml` and `uv.lock`.
+  </Step>
+  <Step title="UV resolves dependencies">
+    UV reads `[tool.uv.sources]` to determine which index each package should come from.
+  </Step>
+  <Step title="UV authenticates">
+    For each private index, UV looks up `UV_INDEX_{NAME}_USERNAME` and `UV_INDEX_{NAME}_PASSWORD`
+    from the environment variables you configured in AMP.
+  </Step>
+  <Step title="Packages install">
+    UV downloads and installs all packages — both public (from PyPI) and private (from your registry).
+  </Step>
+  <Step title="Automation runs">
+    Your crew or flow starts with all dependencies available.
+  </Step>
+</Steps>
+
+## Troubleshooting
+
+### Authentication Errors During Build
+
+**Symptom**: Build fails with `401 Unauthorized` or `403 Forbidden` when resolving a private package.
+
+**Check**:
+- The `UV_INDEX_*` environment variable names match your index name exactly (uppercased, hyphens → underscores)
+- Credentials are set in AMP environment variables, not just in a local `.env`
+- Your token/PAT has the required read permissions for the package feed
+- The token hasn't expired (especially relevant for AWS CodeArtifact)
+
+### Package Not Found
+
+**Symptom**: `No matching distribution found for my-private-package`.
+
+**Check**:
+- The index URL in `pyproject.toml` ends with `/simple/`
+- The `[tool.uv.sources]` entry maps the correct package name to the correct index name
+- The package is actually published to your private registry
+- Run `uv lock` locally with the same credentials to verify resolution works
+
+### Lock File Conflicts
+
+**Symptom**: `uv lock` fails or produces unexpected results after adding a private index.
+
+**Solution**: Set the credentials locally and regenerate:
+
+```bash
+export UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+export UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat
+uv lock
+```
+
+Then commit the updated `uv.lock`.
+
+## Related Guides
+
+<CardGroup cols={3}>
+  <Card title="Prepare for Deployment" icon="clipboard-check" href="/en/enterprise/guides/prepare-for-deployment">
+    Verify project structure and dependencies before deploying.
+  </Card>
+  <Card title="Deploy to AMP" icon="rocket" href="/en/enterprise/guides/deploy-to-amp">
+    Deploy your crew or flow and configure environment variables.
+  </Card>
+  <Card title="Update Your Crew" icon="arrows-rotate" href="/en/enterprise/guides/update-crew">
+    Update environment variables and push changes to a running deployment.
+  </Card>
+</CardGroup>

docs/en/guides/migration/migrating-from-langgraph.mdx

@@ -0,0 +1,518 @@
+---
+title: "Moving from LangGraph to CrewAI: A Practical Guide for Engineers"
+description: If you already have built with LangGraph, learn how to quickly port your projects to CrewAI
+icon: switch
+mode: "wide"
+---
+
+You've built agents with LangGraph. You've wrestled with `StateGraph`, wired up conditional edges, and debugged state dictionaries at 2 AM. It works — but somewhere along the way, you started wondering if there's a better path to production.
+
+There is. **CrewAI Flows** gives you the same power — event-driven orchestration, conditional routing, shared state — with dramatically less boilerplate and a mental model that maps cleanly to how you actually think about multi-step AI workflows.
+
+This article walks through the core concepts side by side, shows real code comparisons, and demonstrates why CrewAI Flows is the framework you'll want to reach for next.
+
+---
+
+## The Mental Model Shift
+
+LangGraph asks you to think in **graphs**: nodes, edges, and state dictionaries. Every workflow is a directed graph where you explicitly wire transitions between computation steps. It's powerful, but the abstraction carries overhead — especially when your workflow is fundamentally sequential with a few decision points.
+
+CrewAI Flows asks you to think in **events**: methods that start things, methods that listen for results, and methods that route execution. The topology of your workflow emerges from decorator annotations rather than explicit graph construction. This isn't just syntactic sugar — it changes how you design, read, and maintain your pipelines.
+
+Here's the core mapping:
+
+| LangGraph Concept | CrewAI Flows Equivalent |
+| --- | --- |
+| `StateGraph` class | `Flow` class |
+| `add_node()` | Methods decorated with `@start`, `@listen` |
+| `add_edge()` / `add_conditional_edges()` | `@listen()` / `@router()` decorators |
+| `TypedDict` state | Pydantic `BaseModel` state |
+| `START` / `END` constants | `@start()` decorator / natural method return |
+| `graph.compile()` | `flow.kickoff()` |
+| Checkpointer / persistence | Built-in memory (LanceDB-backed) |
+
+Let's see what this looks like in practice.
+
+---
+
+## Demo 1: A Simple Sequential Pipeline
+
+Imagine you're building a pipeline that takes a topic, researches it, writes a summary, and formats the output. Here's how each framework handles it.
+
+### LangGraph Approach
+
+```python
+from typing import TypedDict
+from langgraph.graph import StateGraph, START, END
+
+class ResearchState(TypedDict):
+    topic: str
+    raw_research: str
+    summary: str
+    formatted_output: str
+
+def research_topic(state: ResearchState) -> dict:
+    # Call an LLM or search API
+    result = llm.invoke(f"Research the topic: {state['topic']}")
+    return {"raw_research": result}
+
+def write_summary(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Summarize this research:\n{state['raw_research']}"
+    )
+    return {"summary": result}
+
+def format_output(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Format this summary as a polished article section:\n{state['summary']}"
+    )
+    return {"formatted_output": result}
+
+# Build the graph
+graph = StateGraph(ResearchState)
+graph.add_node("research", research_topic)
+graph.add_node("summarize", write_summary)
+graph.add_node("format", format_output)
+
+graph.add_edge(START, "research")
+graph.add_edge("research", "summarize")
+graph.add_edge("summarize", "format")
+graph.add_edge("format", END)
+
+# Compile and run
+app = graph.compile()
+result = app.invoke({"topic": "quantum computing advances in 2026"})
+print(result["formatted_output"])
+```
+
+You define functions, register them as nodes, and manually wire every transition. For a simple sequence like this, there's a lot of ceremony.
+
+### CrewAI Flows Approach
+
+```python
+from crewai import LLM, Agent, Crew, Process, Task
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ResearchState(BaseModel):
+    topic: str = ""
+    raw_research: str = ""
+    summary: str = ""
+    formatted_output: str = ""
+
+class ResearchFlow(Flow[ResearchState]):
+    @start()
+    def research_topic(self):
+        # Option 1: Direct LLM call
+        result = llm.call(f"Research the topic: {self.state.topic}")
+        self.state.raw_research = result
+        return result
+
+    @listen(research_topic)
+    def write_summary(self, research_output):
+        # Option 2: A single agent
+        summarizer = Agent(
+            role="Research Summarizer",
+            goal="Produce concise, accurate summaries of research content",
+            backstory="You are an expert at distilling complex research into clear, "
+            "digestible summaries.",
+            llm=llm,
+            verbose=True,
+        )
+        result = summarizer.kickoff(
+            f"Summarize this research:\n{self.state.raw_research}"
+        )
+        self.state.summary = str(result)
+        return self.state.summary
+
+    @listen(write_summary)
+    def format_output(self, summary_output):
+        # Option 3: a complete crew (with one or more agents)
+        formatter = Agent(
+            role="Content Formatter",
+            goal="Transform research summaries into polished, publication-ready article sections",
+            backstory="You are a skilled editor with expertise in structuring and "
+            "presenting technical content for a general audience.",
+            llm=llm,
+            verbose=True,
+        )
+        format_task = Task(
+            description=f"Format this summary as a polished article section:\n{self.state.summary}",
+            expected_output="A well-structured, polished article section ready for publication.",
+            agent=formatter,
+        )
+        crew = Crew(
+            agents=[formatter],
+            tasks=[format_task],
+            process=Process.sequential,
+            verbose=True,
+        )
+        result = crew.kickoff()
+        self.state.formatted_output = str(result)
+        return self.state.formatted_output
+
+# Run the flow
+flow = ResearchFlow()
+flow.state.topic = "quantum computing advances in 2026"
+result = flow.kickoff()
+print(flow.state.formatted_output)
+
+```
+
+Notice what's different: no graph construction, no edge wiring, no compile step. The execution order is declared right where the logic lives. `@start()` marks the entry point, and `@listen(method_name)` chains steps together. The state is a proper Pydantic model with type safety, validation, and IDE auto-completion.
+
+---
+
+## Demo 2: Conditional Routing
+
+This is where things get interesting. Say you're building a content pipeline that routes to different processing paths based on the type of content detected.
+
+### LangGraph Approach
+
+```python
+from typing import TypedDict, Literal
+from langgraph.graph import StateGraph, START, END
+
+class ContentState(TypedDict):
+    input_text: str
+    content_type: str
+    result: str
+
+def classify_content(state: ContentState) -> dict:
+    content_type = llm.invoke(
+        f"Classify this content as 'technical', 'creative', or 'business':\n{state['input_text']}"
+    )
+    return {"content_type": content_type.strip().lower()}
+
+def process_technical(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as technical doc:\n{state['input_text']}")
+    return {"result": result}
+
+def process_creative(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as creative writing:\n{state['input_text']}")
+    return {"result": result}
+
+def process_business(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as business content:\n{state['input_text']}")
+    return {"result": result}
+
+# Routing function
+def route_content(state: ContentState) -> Literal["technical", "creative", "business"]:
+    return state["content_type"]
+
+# Build the graph
+graph = StateGraph(ContentState)
+graph.add_node("classify", classify_content)
+graph.add_node("technical", process_technical)
+graph.add_node("creative", process_creative)
+graph.add_node("business", process_business)
+
+graph.add_edge(START, "classify")
+graph.add_conditional_edges(
+    "classify",
+    route_content,
+    {
+        "technical": "technical",
+        "creative": "creative",
+        "business": "business",
+    }
+)
+graph.add_edge("technical", END)
+graph.add_edge("creative", END)
+graph.add_edge("business", END)
+
+app = graph.compile()
+result = app.invoke({"input_text": "Explain how TCP handshakes work"})
+```
+
+You need a separate routing function, explicit conditional edge mapping, and termination edges for every branch. The routing logic is decoupled from the node that produces the routing decision.
+
+### CrewAI Flows Approach
+
+```python
+from crewai import LLM, Agent
+from crewai.flow.flow import Flow, listen, router, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ContentState(BaseModel):
+    input_text: str = ""
+    content_type: str = ""
+    result: str = ""
+
+class ContentFlow(Flow[ContentState]):
+    @start()
+    def classify_content(self):
+        self.state.content_type = (
+            llm.call(
+                f"Classify this content as 'technical', 'creative', or 'business':\n"
+                f"{self.state.input_text}"
+            )
+            .strip()
+            .lower()
+        )
+        return self.state.content_type
+
+    @router(classify_content)
+    def route_content(self, classification):
+        if classification == "technical":
+            return "process_technical"
+        elif classification == "creative":
+            return "process_creative"
+        else:
+            return "process_business"
+
+    @listen("process_technical")
+    def handle_technical(self):
+        agent = Agent(
+            role="Technical Writer",
+            goal="Produce clear, accurate technical documentation",
+            backstory="You are an expert technical writer who specializes in "
+            "explaining complex technical concepts precisely.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as technical doc:\n{self.state.input_text}")
+        )
+
+    @listen("process_creative")
+    def handle_creative(self):
+        agent = Agent(
+            role="Creative Writer",
+            goal="Craft engaging and imaginative creative content",
+            backstory="You are a talented creative writer with a flair for "
+            "compelling storytelling and vivid expression.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as creative writing:\n{self.state.input_text}")
+        )
+
+    @listen("process_business")
+    def handle_business(self):
+        agent = Agent(
+            role="Business Writer",
+            goal="Produce professional, results-oriented business content",
+            backstory="You are an experienced business writer who communicates "
+            "strategy and value clearly to professional audiences.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as business content:\n{self.state.input_text}")
+        )
+
+flow = ContentFlow()
+flow.state.input_text = "Explain how TCP handshakes work"
+flow.kickoff()
+print(flow.state.result)
+
+```
+
+The `@router()` decorator turns a method into a decision point. It returns a string that matches a listener — no mapping dictionaries, no separate routing functions. The branching logic reads like a Python `if` statement because it *is* one.
+
+---
+
+## Demo 3: Integrating AI Agent Crews into Flows
+
+Here's where CrewAI's real power shines. Flows aren't just for chaining LLM calls — they orchestrate full **Crews** of autonomous agents. This is something LangGraph simply doesn't have a native equivalent for.
+
+```python
+from crewai import Agent, Task, Crew
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+class ArticleState(BaseModel):
+    topic: str = ""
+    research: str = ""
+    draft: str = ""
+    final_article: str = ""
+
+class ArticleFlow(Flow[ArticleState]):
+
+    @start()
+    def run_research_crew(self):
+        """A full Crew of agents handles research."""
+        researcher = Agent(
+            role="Senior Research Analyst",
+            goal=f"Produce comprehensive research on: {self.state.topic}",
+            backstory="You're a veteran analyst known for thorough, "
+                       "well-sourced research reports.",
+            llm="gpt-4o"
+        )
+
+        research_task = Task(
+            description=f"Research '{self.state.topic}' thoroughly. "
+                        "Cover key trends, data points, and expert opinions.",
+            expected_output="A detailed research brief with sources.",
+            agent=researcher
+        )
+
+        crew = Crew(agents=[researcher], tasks=[research_task])
+        result = crew.kickoff()
+        self.state.research = result.raw
+        return result.raw
+
+    @listen(run_research_crew)
+    def run_writing_crew(self, research_output):
+        """A different Crew handles writing."""
+        writer = Agent(
+            role="Technical Writer",
+            goal="Write a compelling article based on provided research.",
+            backstory="You turn complex research into engaging, clear prose.",
+            llm="gpt-4o"
+        )
+
+        editor = Agent(
+            role="Senior Editor",
+            goal="Review and polish articles for publication quality.",
+            backstory="20 years of editorial experience at top tech publications.",
+            llm="gpt-4o"
+        )
+
+        write_task = Task(
+            description=f"Write an article based on this research:\n{self.state.research}",
+            expected_output="A well-structured draft article.",
+            agent=writer
+        )
+
+        edit_task = Task(
+            description="Review, fact-check, and polish the draft article.",
+            expected_output="A publication-ready article.",
+            agent=editor
+        )
+
+        crew = Crew(agents=[writer, editor], tasks=[write_task, edit_task])
+        result = crew.kickoff()
+        self.state.final_article = result.raw
+        return result.raw
+
+# Run the full pipeline
+flow = ArticleFlow()
+flow.state.topic = "The Future of Edge AI"
+flow.kickoff()
+print(flow.state.final_article)
+```
+
+This is the key insight: **Flows provide the orchestration layer, and Crews provide the intelligence layer.** Each step in a Flow can spin up a full team of collaborating agents, each with their own roles, goals, and tools. You get structured, predictable control flow *and* autonomous agent collaboration — the best of both worlds.
+
+In LangGraph, achieving something similar means manually implementing agent communication protocols, tool-calling loops, and delegation logic inside your node functions. It's possible, but it's plumbing you're building from scratch every time.
+
+---
+
+## Demo 4: Parallel Execution and Synchronization
+
+Real-world pipelines often need to fan out work and join the results. CrewAI Flows handles this elegantly with `and_` and `or_` operators.
+
+```python
+from crewai import LLM
+from crewai.flow.flow import Flow, and_, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class AnalysisState(BaseModel):
+    topic: str = ""
+    market_data: str = ""
+    tech_analysis: str = ""
+    competitor_intel: str = ""
+    final_report: str = ""
+
+class ParallelAnalysisFlow(Flow[AnalysisState]):
+    @start()
+    def start_method(self):
+        pass
+
+    @listen(start_method)
+    def gather_market_data(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def run_tech_analysis(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def gather_competitor_intel(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(and_(gather_market_data, run_tech_analysis, gather_competitor_intel))
+    def synthesize_report(self):
+        # Your agentic or deterministic code
+        pass
+
+flow = ParallelAnalysisFlow()
+flow.state.topic = "AI-powered developer tools"
+flow.kickoff()
+
+```
+
+Multiple `@start()` decorators fire in parallel. The `and_()` combinator on the `@listen` decorator ensures `synthesize_report` only executes after *all three* upstream methods complete. There's also `or_()` for when you want to proceed as soon as *any* upstream task finishes.
+
+In LangGraph, you'd need to build a fan-out/fan-in pattern with parallel branches, a synchronization node, and careful state merging — all wired explicitly through edges.
+
+---
+
+## Why CrewAI Flows for Production
+
+Beyond cleaner syntax, Flows deliver several production-critical advantages:
+
+**Built-in state persistence.** Flow state is backed by LanceDB, meaning your workflows can survive crashes, be resumed, and accumulate knowledge across runs. LangGraph requires you to configure a separate checkpointer.
+
+**Type-safe state management.** Pydantic models give you validation, serialization, and IDE support out of the box. LangGraph's `TypedDict` states don't validate at runtime.
+
+**First-class agent orchestration.** Crews are a native primitive. You define agents with roles, goals, backstories, and tools — and they collaborate autonomously within the structured envelope of a Flow. No need to reinvent multi-agent coordination.
+
+**Simpler mental model.** Decorators declare intent. `@start` means "begin here." `@listen(x)` means "run after x." `@router(x)` means "decide where to go after x." The code reads like the workflow it describes.
+
+**CLI integration.** Run flows with `crewai run`. No separate compilation step, no graph serialization. Your Flow is a Python class, and it runs like one.
+
+---
+
+## Migration Cheat Sheet
+
+If you're sitting on a LangGraph codebase and want to move to CrewAI Flows, here's a practical conversion guide:
+
+1. **Map your state.** Convert your `TypedDict` to a Pydantic `BaseModel`. Add default values for all fields.
+2. **Convert nodes to methods.** Each `add_node` function becomes a method on your `Flow` subclass. Replace `state["field"]` reads with `self.state.field`.
+3. **Replace edges with decorators.** Your `add_edge(START, "first_node")` becomes `@start()` on the first method. Sequential `add_edge("a", "b")` becomes `@listen(a)` on method `b`.
+4. **Replace conditional edges with `@router`.** Your routing function and `add_conditional_edges()` mapping become a single `@router()` method that returns a route string.
+5. **Replace compile + invoke with kickoff.** Drop `graph.compile()`. Call `flow.kickoff()` instead.
+6. **Consider where Crews fit.** Any node where you have complex multi-step agent logic is a candidate for extraction into a Crew. This is where you'll see the biggest quality improvement.
+
+---
+
+## Getting Started
+
+Install CrewAI and scaffold a new Flow project:
+
+```bash
+pip install crewai
+crewai create flow my_first_flow
+cd my_first_flow
+```
+
+This generates a project structure with a ready-to-edit Flow class, configuration files, and a `pyproject.toml` with `type = "flow"` already set. Run it with:
+
+```bash
+crewai run
+```
+
+From there, add your agents, wire up your listeners, and ship it.
+
+---
+
+## Final Thoughts
+
+LangGraph taught the ecosystem that AI workflows need structure. That was an important lesson. But CrewAI Flows takes that lesson and delivers it in a form that's faster to write, easier to read, and more powerful in production — especially when your workflows involve multiple collaborating agents.
+
+If you're building anything beyond a single-agent chain, give Flows a serious look. The decorator-driven model, native Crew integration, and built-in state management mean you'll spend less time on plumbing and more time on the problems that matter.
+
+Start with `crewai create flow`. You won't look back.

docs/en/learn/human-feedback-in-flows.mdx

@@ -98,33 +98,43 @@ def handle_feedback(self, result):
 When you specify `emit`, the decorator becomes a router. The human's free-form feedback is interpreted by an LLM and collapsed into one of the specified outcomes:
 
 ```python Code
-@start()
-@human_feedback(
-    message="Do you approve this content for publication?",
-    emit=["approved", "rejected", "needs_revision"],
-    llm="gpt-4o-mini",
-    default_outcome="needs_revision",
-)
-def review_content(self):
-    return "Draft blog post content here..."
+from crewai.flow.flow import Flow, start, listen, or_
+from crewai.flow.human_feedback import human_feedback
 
-@listen("approved")
-def publish(self, result):
-    print(f"Publishing! User said: {result.feedback}")
+class ReviewFlow(Flow):
+    @start()
+    def generate_content(self):
+        return "Draft blog post content here..."
 
-@listen("rejected")
-def discard(self, result):
-    print(f"Discarding. Reason: {result.feedback}")
+    @human_feedback(
+        message="Do you approve this content for publication?",
+        emit=["approved", "rejected", "needs_revision"],
+        llm="gpt-4o-mini",
+        default_outcome="needs_revision",
+    )
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "Draft blog post content here..."
 
-@listen("needs_revision")
-def revise(self, result):
-    print(f"Revising based on: {result.feedback}")
+    @listen("approved")
+    def publish(self, result):
+        print(f"Publishing! User said: {result.feedback}")
+
+    @listen("rejected")
+    def discard(self, result):
+        print(f"Discarding. Reason: {result.feedback}")
 ```
 
+When the human says something like "needs more detail", the LLM collapses that to `"needs_revision"`, which triggers `review_content` again via `or_()` — creating a revision loop. The loop continues until the outcome is `"approved"` or `"rejected"`.
+
 <Tip>
 The LLM uses structured outputs (function calling) when available to guarantee the response is one of your specified outcomes. This makes routing reliable and predictable.
 </Tip>
 
+<Warning>
+A `@start()` method only runs once at the beginning of the flow. If you need a revision loop, separate the start method from the review method and use `@listen(or_("trigger", "revision_outcome"))` on the review method to enable the self-loop.
+</Warning>
+
 ## HumanFeedbackResult
 
 The `HumanFeedbackResult` dataclass contains all information about a human feedback interaction:
@@ -188,127 +198,183 @@ Each `HumanFeedbackResult` is appended to `human_feedback_history`, so multiple
 
 ## Complete Example: Content Approval Workflow
 
-Here's a full example implementing a content review and approval workflow:
+Here's a full example implementing a content review and approval workflow with a revision loop:
 
 <CodeGroup>
 
 ```python Code
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 from pydantic import BaseModel
 
 
 class ContentState(BaseModel):
-    topic: str = ""
     draft: str = ""
-    final_content: str = ""
     revision_count: int = 0
+    status: str = "pending"
 
 
 class ContentApprovalFlow(Flow[ContentState]):
-    """A flow that generates content and gets human approval."""
+    """A flow that generates content and loops until the human approves."""
 
     @start()
-    def get_topic(self):
-        self.state.topic = input("What topic should I write about? ")
-        return self.state.topic
-
-    @listen(get_topic)
-    def generate_draft(self, topic):
-        # In real use, this would call an LLM
-        self.state.draft = f"# {topic}\n\nThis is a draft about {topic}..."
+    def generate_draft(self):
+        self.state.draft = "# AI Safety\n\nThis is a draft about AI Safety..."
         return self.state.draft
 
-    @listen(generate_draft)
     @human_feedback(
-        message="Please review this draft. Reply 'approved', 'rejected', or provide revision feedback:",
+        message="Please review this draft. Approve, reject, or describe what needs changing:",
         emit=["approved", "rejected", "needs_revision"],
         llm="gpt-4o-mini",
         default_outcome="needs_revision",
     )
-    def review_draft(self, draft):
-        return draft
+    @listen(or_("generate_draft", "needs_revision"))
+    def review_draft(self):
+        self.state.revision_count += 1
+        return f"{self.state.draft} (v{self.state.revision_count})"
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
-        self.state.final_content = result.output
-        print("\n✅ Content approved and published!")
-        print(f"Reviewer comment: {result.feedback}")
+        self.state.status = "published"
+        print(f"Content approved and published! Reviewer said: {result.feedback}")
         return "published"
 
     @listen("rejected")
     def handle_rejection(self, result: HumanFeedbackResult):
-        print("\n❌ Content rejected")
-        print(f"Reason: {result.feedback}")
+        self.state.status = "rejected"
+        print(f"Content rejected. Reason: {result.feedback}")
         return "rejected"
 
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        self.state.revision_count += 1
-        print(f"\n📝 Revision #{self.state.revision_count} requested")
-        print(f"Feedback: {result.feedback}")
 
-        # In a real flow, you might loop back to generate_draft
-        # For this example, we just acknowledge
-        return "revision_requested"
-
-
-# Run the flow
 flow = ContentApprovalFlow()
 result = flow.kickoff()
-print(f"\nFlow completed. Revisions requested: {flow.state.revision_count}")
+print(f"\nFlow completed. Status: {flow.state.status}, Reviews: {flow.state.revision_count}")
 ```
 
 ```text Output
-What topic should I write about? AI Safety
+==================================================
+OUTPUT FOR REVIEW:
+==================================================
+# AI Safety
+
+This is a draft about AI Safety... (v1)
+==================================================
+
+Please review this draft. Approve, reject, or describe what needs changing:
+(Press Enter to skip, or type your feedback)
+
+Your feedback: Needs more detail on alignment research
 
 ==================================================
 OUTPUT FOR REVIEW:
 ==================================================
 # AI Safety
 
-This is a draft about AI Safety...
+This is a draft about AI Safety... (v2)
 ==================================================
 
-Please review this draft. Reply 'approved', 'rejected', or provide revision feedback:
+Please review this draft. Approve, reject, or describe what needs changing:
 (Press Enter to skip, or type your feedback)
 
 Your feedback: Looks good, approved!
 
-✅ Content approved and published!
-Reviewer comment: Looks good, approved!
+Content approved and published! Reviewer said: Looks good, approved!
 
-Flow completed. Revisions requested: 0
+Flow completed. Status: published, Reviews: 2
 ```
 
 </CodeGroup>
 
+The key pattern is `@listen(or_("generate_draft", "needs_revision"))` — the review method listens to both the initial trigger and its own revision outcome, creating a self-loop that repeats until the human approves or rejects.
+
 ## Combining with Other Decorators
 
-The `@human_feedback` decorator works with other flow decorators. Place it as the innermost decorator (closest to the function):
+The `@human_feedback` decorator works with `@start()`, `@listen()`, and `or_()`. Both decorator orderings work — the framework propagates attributes in both directions — but the recommended patterns are:
 
 ```python Code
-# Correct: @human_feedback is innermost (closest to the function)
+# One-shot review at the start of a flow (no self-loop)
 @start()
-@human_feedback(message="Review this:")
+@human_feedback(message="Review this:", emit=["approved", "rejected"], llm="gpt-4o-mini")
 def my_start_method(self):
     return "content"
 
+# Linear review on a listener (no self-loop)
 @listen(other_method)
-@human_feedback(message="Review this too:")
+@human_feedback(message="Review this too:", emit=["good", "bad"], llm="gpt-4o-mini")
 def my_listener(self, data):
     return f"processed: {data}"
+
+# Self-loop: review that can loop back for revisions
+@human_feedback(message="Approve or revise?", emit=["approved", "revise"], llm="gpt-4o-mini")
+@listen(or_("upstream_method", "revise"))
+def review_with_loop(self):
+    return "content for review"
 ```
 
-<Tip>
-Place `@human_feedback` as the innermost decorator (last/closest to the function) so it wraps the method directly and can capture the return value before passing to the flow system.
-</Tip>
+### Self-loop pattern
+
+To create a revision loop, the review method must listen to **both** an upstream trigger and its own revision outcome using `or_()`:
+
+```python Code
+@start()
+def generate(self):
+    return "initial draft"
+
+@human_feedback(
+    message="Approve or request changes?",
+    emit=["revise", "approved"],
+    llm="gpt-4o-mini",
+    default_outcome="approved",
+)
+@listen(or_("generate", "revise"))
+def review(self):
+    return "content"
+
+@listen("approved")
+def publish(self):
+    return "published"
+```
+
+When the outcome is `"revise"`, the flow routes back to `review` (because it listens to `"revise"` via `or_()`). When the outcome is `"approved"`, the flow continues to `publish`. This works because the flow engine exempts routers from the "fire once" rule, allowing them to re-execute on each loop iteration.
+
+### Chained routers
+
+A listener triggered by one router's outcome can itself be a router:
+
+```python Code
+@start()
+def generate(self):
+    return "draft content"
+
+@human_feedback(message="First review:", emit=["approved", "rejected"], llm="gpt-4o-mini")
+@listen("generate")
+def first_review(self):
+    return "draft content"
+
+@human_feedback(message="Final review:", emit=["publish", "hold"], llm="gpt-4o-mini")
+@listen("approved")
+def final_review(self, prev):
+    return "final content"
+
+@listen("publish")
+def on_publish(self, prev):
+    return "published"
+
+@listen("hold")
+def on_hold(self, prev):
+    return "held for later"
+```
+
+### Limitations
+
+- **`@start()` methods run once**: A `@start()` method cannot self-loop. If you need a revision cycle, use a separate `@start()` method as the entry point and put the `@human_feedback` on a `@listen()` method.
+- **No `@start()` + `@listen()` on the same method**: This is a Flow framework constraint. A method is either a start point or a listener, not both.
 
 ## Best Practices
 
 ### 1. Write Clear Request Messages
 
-The `request` parameter is what the human sees. Make it actionable:
+The `message` parameter is what the human sees. Make it actionable:
 
 ```python Code
 # ✅ Good - clear and actionable
@@ -516,9 +582,9 @@ class ContentPipeline(Flow):
     @start()
     @human_feedback(
         message="Approve this content for publication?",
-        emit=["approved", "rejected", "needs_revision"],
+        emit=["approved", "rejected"],
         llm="gpt-4o-mini",
-        default_outcome="needs_revision",
+        default_outcome="rejected",
         provider=SlackNotificationProvider("#content-reviews"),
     )
     def generate_content(self):
@@ -534,11 +600,6 @@ class ContentPipeline(Flow):
         print(f"Archived. Reason: {result.feedback}")
         return {"status": "archived"}
 
-    @listen("needs_revision")
-    def queue_revision(self, result):
-        print(f"Queued for revision: {result.feedback}")
-        return {"status": "revision_needed"}
-
 
 # Starting the flow (will pause and wait for Slack response)
 def start_content_pipeline():
@@ -594,22 +655,22 @@ Over time, the human sees progressively better pre-reviewed output because each
 ```python Code
 class ArticleReviewFlow(Flow):
     @start()
+    def generate_article(self):
+        return self.crew.kickoff(inputs={"topic": "AI Safety"}).raw
+
     @human_feedback(
         message="Review this article draft:",
         emit=["approved", "needs_revision"],
         llm="gpt-4o-mini",
         learn=True,  # enable HITL learning
     )
-    def generate_article(self):
-        return self.crew.kickoff(inputs={"topic": "AI Safety"}).raw
+    @listen(or_("generate_article", "needs_revision"))
+    def review_article(self):
+        return self.last_human_feedback.output if self.last_human_feedback else "article draft"
 
     @listen("approved")
     def publish(self):
         print(f"Publishing: {self.last_human_feedback.output}")
-
-    @listen("needs_revision")
-    def revise(self):
-        print("Revising based on feedback...")
 ```
 
 **First run**: The human sees the raw output and says "Always include citations for factual claims." The lesson is distilled and stored in memory.

docs/en/learn/llm-connections.mdx

@@ -7,7 +7,7 @@ mode: "wide"
 
 ## Connect CrewAI to LLMs
 
-CrewAI uses LiteLLM to connect to a wide variety of Language Models (LLMs). This integration provides extensive versatility, allowing you to use models from numerous providers with a simple, unified interface.
+CrewAI connects to LLMs through native SDK integrations for the most popular providers (OpenAI, Anthropic, Google Gemini, Azure, and AWS Bedrock), and uses LiteLLM as a flexible fallback for all other providers.
 
 <Note>
     By default, CrewAI uses the `gpt-4o-mini` model. This is determined by the `OPENAI_MODEL_NAME` environment variable, which defaults to "gpt-4o-mini" if not set.
@@ -41,6 +41,14 @@ LiteLLM supports a wide range of providers, including but not limited to:
 
 For a complete and up-to-date list of supported providers, please refer to the [LiteLLM Providers documentation](https://docs.litellm.ai/docs/providers).
 
+<Info>
+  To use any provider not covered by a native integration, add LiteLLM as a dependency to your project:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+  Native providers (OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock) use their own SDK extras — see the [Provider Configuration Examples](/en/concepts/llms#provider-configuration-examples).
+</Info>
+
 ## Changing the LLM
 
 To use a different LLM with your CrewAI agents, you have several options:

docs/en/observability/tracing.mdx

@@ -35,7 +35,7 @@ Visit [app.crewai.com](https://app.crewai.com) and create your free account. Thi
 If you haven't already, install CrewAI with the CLI tools:
 
 ```bash
-uv add crewai[tools]
+uv add 'crewai[tools]'
 ```
 
 Then authenticate your CLI with your CrewAI AMP account:

docs/en/tools/automation/composiotool.mdx

@@ -18,77 +18,46 @@ Composio is an integration platform that allows you to connect your AI agents to
 To incorporate Composio tools into your project, follow the instructions below:
 
 ```shell
-pip install composio-crewai
+pip install composio composio-crewai
 pip install crewai
 ```
 
-After the installation is complete, either run `composio login` or export your composio API key as `COMPOSIO_API_KEY`. Get your Composio API key from [here](https://app.composio.dev)
+After the installation is complete, set your Composio API key as `COMPOSIO_API_KEY`. Get your Composio API key from [here](https://platform.composio.dev)
 
 ## Example
 
 The following example demonstrates how to initialize the tool and execute a github action:
 
-1. Initialize Composio toolset
+1. Initialize Composio with CrewAI Provider
 
 ```python Code
-from composio_crewai import ComposioToolSet, App, Action
+from composio_crewai import ComposioProvider
+from composio import Composio
 from crewai import Agent, Task, Crew
 
-toolset = ComposioToolSet()
+composio = Composio(provider=ComposioProvider())
 ```
 
-2. Connect your GitHub account
+2. Create a new Composio Session and retrieve the tools
 <CodeGroup>
-```shell CLI
-composio add github
-```
-```python Code
-request = toolset.initiate_connection(app=App.GITHUB)
-print(f"Open this URL to authenticate: {request.redirectUrl}")
+```python
+session = composio.create(
+    user_id="your-user-id",
+    toolkits=["gmail", "github"] # optional, default is all toolkits
+)
+tools = session.tools()
 ```
+Read more about sessions and user management [here](https://docs.composio.dev/docs/configuring-sessions)
 </CodeGroup>
 
-3. Get Tools
+3. Authenticating users manually
 
-- Retrieving all the tools from an app (not recommended for production):
+Composio automatically authenticates the users during the agent chat session. However, you can also authenticate the user manually by calling the `authorize` method.
 ```python Code
-tools = toolset.get_tools(apps=[App.GITHUB])
+connection_request = session.authorize("github")
+print(f"Open this URL to authenticate: {connection_request.redirect_url}")
 ```
 
-- Filtering tools based on tags:
-```python Code
-tag = "users"
-
-filtered_action_enums = toolset.find_actions_by_tags(
-    App.GITHUB,
-    tags=[tag], 
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-
-- Filtering tools based on use case:
-```python Code
-use_case = "Star a repository on GitHub"
-
-filtered_action_enums = toolset.find_actions_by_use_case(
-    App.GITHUB, use_case=use_case, advanced=False
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-<Tip>Set `advanced` to True to get actions for complex use cases</Tip>
-
-- Using specific tools:
-
-In this demo, we will use the `GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER` action from the GitHub app.
-```python Code
-tools = toolset.get_tools(
-    actions=[Action.GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER]
-)
-```
-Learn more about filtering actions [here](https://docs.composio.dev/patterns/tools/use-tools/use-specific-actions)
-
 4. Define agent
 
 ```python Code
@@ -116,4 +85,4 @@ crew = Crew(agents=[crewai_agent], tasks=[task])
 crew.kickoff()
 ```
 
-* More detailed list of tools can be found [here](https://app.composio.dev)
+* More detailed list of tools can be found [here](https://docs.composio.dev/toolkits)

docs/ko/changelog.mdx

@@ -4,6 +4,138 @@ description: "CrewAI의 제품 업데이트, 개선 사항 및 버그 수정"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="2026년 3월 4일">
+  ## v1.10.1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## 변경 사항
+
+  ### 기능
+  - Gemini GenAI 업그레이드
+
+  ### 버그 수정
+  - 재귀를 피하기 위해 실행기 리스너 값을 조정
+  - Gemini에서 병렬 함수 응답 부분을 단일 Content 객체로 그룹화
+  - Gemini에서 사고 모델의 사고 출력을 표시
+  - 에이전트 도구가 None일 때 MCP 및 플랫폼 도구 로드
+  - A2A에서 실행 이벤트 루프가 있는 Jupyter 환경 지원
+  - 일시적인 추적을 위해 익명 ID 사용
+  - 조건부로 플러스 헤더 전달
+  - 원격 측정을 위해 비주 스레드에서 신호 처리기 등록 건너뛰기
+  - 도구 오류를 관찰로 주입하고 이름 충돌 해결
+  - Dependabot 경고를 해결하기 위해 pypdf를 4.x에서 6.7.4로 업그레이드
+  - 심각 및 높은 Dependabot 보안 경고 해결
+
+  ### 문서
+  - Composio 도구 문서를 지역별로 동기화
+
+  ## 기여자
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
+<Update label="2026년 2월 27일">
+  ## v1.10.1a1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## 변경 사항
+
+  ### 기능
+  - 단계 콜백 메서드에서 비동기 호출 지원 구현
+  - 메모리 모듈의 무거운 의존성에 대한 지연 로딩 구현
+
+  ### 문서
+  - v1.10.0에 대한 변경 로그 및 버전 업데이트
+
+  ### 리팩토링
+  - 비동기 호출을 지원하기 위해 단계 콜백 메서드 리팩토링
+  - 메모리 모듈의 무거운 의존성에 대한 지연 로딩을 구현하기 위해 리팩토링
+
+  ### 버그 수정
+  - 릴리스 노트의 분기 수정
+
+  ## 기여자
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="2026년 2월 27일">
+  ## v1.10.1a1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## 변경 사항
+
+  ### 리팩토링
+  - 비동기 호출을 지원하기 위해 단계 콜백 메서드 리팩토링
+  - 메모리 모듈의 무거운 의존성에 대해 지연 로딩 구현
+
+  ### 문서화
+  - v1.10.0에 대한 변경 로그 및 버전 업데이트
+
+  ### 버그 수정
+  - 릴리스 노트를 위한 브랜치 생성
+
+  ## 기여자
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="2026년 2월 26일">
+  ## v1.10.0
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.10.0)
+
+  ## 변경 사항
+
+  ### 기능
+  - MCP 도구 해상도 및 관련 이벤트 개선
+  - lancedb 버전 업데이트 및 lance-namespace 패키지 추가
+  - CrewAgentExecutor 및 BaseTool에서 JSON 인수 파싱 및 검증 개선
+  - CLI HTTP 클라이언트를 requests에서 httpx로 마이그레이션
+  - 버전화된 문서 추가
+  - 버전 노트에 대한 yanked 감지 추가
+  - Flows에서 사용자 입력 처리 구현
+  - 인간 피드백 통합 테스트에서 HITL 자기 루프 기능 개선
+  - eventbus에 started_event_id 추가 및 설정
+  - tools.specs 자동 업데이트
+
+  ### 버그 수정
+  - 빈 경우에도 도구 kwargs를 검증하여 모호한 TypeError 방지
+  - LLM을 위한 도구 매개변수 스키마에서 null 타입 유지
+  - output_pydantic/output_json을 네이티브 구조화된 출력으로 매핑
+  - 약속이 있는 경우 콜백이 실행/대기되도록 보장
+  - 예외 컨텍스트에서 메서드 이름 캡처
+  - 라우터 결과에서 enum 타입 유지; 타입 개선
+  - 입력으로 지속성 ID가 전달될 때 조용히 깨지는 순환 흐름 수정
+  - CLI 플래그 형식을 --skip-provider에서 --skip_provider로 수정
+  - OpenAI 도구 호출 스트림이 완료되도록 보장
+  - MCP 도구에서 복잡한 스키마 $ref 포인터 해결
+  - 스키마에서 additionalProperties=false 강제 적용
+  - 크루 폴더에 대해 예약된 스크립트 이름 거부
+  - 가드레일 이벤트 방출 테스트에서 경쟁 조건 해결
+
+  ### 문서
+  - 비네이티브 LLM 공급자를 위한 litellm 종속성 노트 추가
+  - NL2SQL 보안 모델 및 강화 지침 명확화
+  - 9개 통합에서 96개의 누락된 작업 추가
+
+  ### 리팩토링
+  - crew를 provider로 리팩토링
+  - HITL을 provider 패턴으로 추출
+  - 훅 타이핑 및 등록 개선
+
+  ## 기여자
+
+  @dependabot[bot], @github-actions[bot], @github-code-quality[bot], @greysonlalonde, @heitorado, @hobostay, @joaomdmoura, @johnvan7, @jonathansampson, @lorenzejay, @lucasgomide, @mattatcha, @mplachta, @nicoferdi96, @theCyberTech, @thiagomoretto, @vinibrsl
+
+</Update>
+
 <Update label="2026년 1월 26일">
   ## v1.9.0
 

docs/ko/concepts/llms.mdx

@@ -105,6 +105,15 @@ CrewAI 코드 내에는 사용할 모델을 지정할 수 있는 여러 위치
   </Tab>
 </Tabs>
 
+<Info>
+  CrewAI는 OpenAI, Anthropic, Google (Gemini API), Azure, AWS Bedrock에 대해 네이티브 SDK 통합을 제공합니다 — 제공자별 extras(예: `uv add "crewai[openai]"`) 외에 추가 설치가 필요하지 않습니다.
+
+  그 외 모든 제공자는 **LiteLLM**을 통해 지원됩니다. 이를 사용하려면 프로젝트에 의존성으로 추가하세요:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+</Info>
+
 ## 공급자 구성 예시
 
 CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양한 LLM 공급자를 지원합니다.
@@ -214,6 +223,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
     | `meta_llama/Llama-4-Maverick-17B-128E-Instruct-FP8` | 128k | 4028 | 텍스트, 이미지     | 텍스트           |
     | `meta_llama/Llama-3.3-70B-Instruct`               | 128k | 4028       | 텍스트            | 텍스트           |
     | `meta_llama/Llama-3.3-8B-Instruct`                | 128k | 4028       | 텍스트            | 텍스트           |
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Anthropic">
@@ -354,6 +368,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
     | gemini-1.5-flash                 | 1M 토큰         | 밸런스 잡힌 멀티모달 모델, 대부분의 작업에 적합                         |
     | gemini-1.5-flash-8B              | 1M 토큰         | 가장 빠르고, 비용 효율적, 고빈도 작업에 적합                            |
     | gemini-1.5-pro                   | 2M 토큰         | 최고의 성능, 논리적 추론, 코딩, 창의적 협업 등 다양한 추론 작업에 적합   |
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Azure">
@@ -439,6 +458,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         model="sagemaker/<my-endpoint>"
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Mistral">
@@ -454,6 +478,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         temperature=0.7
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Nvidia NIM">
@@ -540,6 +569,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
     | rakuten/rakutenai-7b-instruct                                          | 1,024 토큰     | 언어 이해, 추론, 텍스트 생성이 탁월한 최첨단 LLM                         |
     | rakuten/rakutenai-7b-chat                                              | 1,024 토큰     | 언어 이해, 추론, 텍스트 생성이 탁월한 최첨단 LLM                         |
     | baichuan-inc/baichuan2-13b-chat                                        | 4,096 토큰     | 중국어 및 영어 대화, 코딩, 수학, 지시 따르기, 퀴즈 풀이 지원             |
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Local NVIDIA NIM Deployed using WSL2">
@@ -580,6 +614,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
 
         # ...
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Groq">
@@ -601,6 +640,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
     | Llama 3.1 70B/8B| 131,072 토큰      | 고성능, 대용량 문맥 작업         |
     | Llama 3.2 Series| 8,192 토큰        | 범용 작업                        |
     | Mixtral 8x7B    | 32,768 토큰       | 성능과 문맥의 균형               |
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="IBM watsonx.ai">
@@ -623,6 +667,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         base_url="https://api.watsonx.ai/v1"
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Ollama (Local LLMs)">
@@ -636,6 +685,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         base_url="http://localhost:11434"
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Fireworks AI">
@@ -651,6 +705,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         temperature=0.7
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Perplexity AI">
@@ -666,6 +725,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         base_url="https://api.perplexity.ai/"
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Hugging Face">
@@ -680,6 +744,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
         model="huggingface/meta-llama/Meta-Llama-3.1-8B-Instruct"
     )
     ```
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="SambaNova">
@@ -703,6 +772,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
     | Llama 3.2 Series| 8,192 토큰         | 범용, 멀티모달 작업                  |
     | Llama 3.3 70B   | 최대 131,072 토큰   | 고성능, 높은 출력 품질               |
     | Qwen2 familly   | 8,192 토큰         | 고성능, 높은 출력 품질               |
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Cerebras">
@@ -728,6 +802,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
       - 속도와 품질의 우수한 밸런스
       - 긴 컨텍스트 윈도우 지원
     </Info>
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Open Router">
@@ -750,6 +829,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
       - openrouter/deepseek/deepseek-r1
       - openrouter/deepseek/deepseek-chat
     </Info>
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Nebius AI Studio">
@@ -772,6 +856,11 @@ CrewAI는 고유한 기능, 인증 방법, 모델 역량을 제공하는 다양
       - 경쟁력 있는 가격
       - 속도와 품질의 우수한 밸런스
     </Info>
+
+    **참고:** 이 제공자는 LiteLLM을 사용합니다. 프로젝트에 의존성으로 추가하세요:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 </AccordionGroup>
 

docs/ko/enterprise/features/flow-hitl-management.mdx

@@ -38,22 +38,21 @@ CrewAI Enterprise는 AI 워크플로우를 협업적인 인간-AI 프로세스
 `@human_feedback` 데코레이터를 사용하여 Flow 내에 인간 검토 체크포인트를 구성합니다. 실행이 검토 포인트에 도달하면 시스템이 일시 중지되고, 담당자에게 이메일로 알리며, 응답을 기다립니다.
 
 ```python
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 
 class ContentApprovalFlow(Flow):
     @start()
     def generate_content(self):
-        # AI가 콘텐츠 생성
         return "Q1 캠페인용 마케팅 카피 생성..."
 
-    @listen(generate_content)
     @human_feedback(
         message="브랜드 준수를 위해 이 콘텐츠를 검토해 주세요:",
         emit=["approved", "rejected", "needs_revision"],
     )
-    def review_content(self, content):
-        return content
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "검토용 마케팅 카피..."
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
@@ -62,10 +61,6 @@ class ContentApprovalFlow(Flow):
     @listen("rejected")
     def archive_content(self, result: HumanFeedbackResult):
         print(f"콘텐츠 거부됨. 사유: {result.feedback}")
-
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        print(f"수정 요청: {result.feedback}")
 ```
 
 완전한 구현 세부 사항은 [Flow에서 인간 피드백](/ko/learn/human-feedback-in-flows) 가이드를 참조하세요.

docs/ko/enterprise/guides/deploy-to-amp.mdx

@@ -176,6 +176,11 @@ Crew를 GitHub 저장소에 푸시해야 합니다. 아직 Crew를 만들지 않
       ![Set Environment Variables](/images/enterprise/set-env-variables.png)
     </Frame>
 
+    <Info>
+      프라이빗 Python 패키지를 사용하시나요? 여기에 레지스트리 자격 증명도 추가해야 합니다.
+      필요한 변수는 [프라이빗 패키지 레지스트리](/ko/enterprise/guides/private-package-registry)를 참조하세요.
+    </Info>
+
   </Step>
 
   <Step title="Crew 배포하기">

docs/ko/enterprise/guides/prepare-for-deployment.mdx

@@ -256,6 +256,12 @@ Crews와 Flows 모두 `src/project_name/main.py`에 진입점이 있습니다:
 1. **LLM API 키** (OpenAI, Anthropic, Google 등)
 2. **도구 API 키** - 외부 도구를 사용하는 경우 (Serper 등)
 
+<Info>
+  프로젝트가 **프라이빗 PyPI 레지스트리**의 패키지에 의존하는 경우, 레지스트리 인증 자격 증명도
+  환경 변수로 구성해야 합니다. 자세한 내용은
+  [프라이빗 패키지 레지스트리](/ko/enterprise/guides/private-package-registry) 가이드를 참조하세요.
+</Info>
+
 <Tip>
   구성 문제를 조기에 발견하기 위해 배포 전에 동일한 환경 변수로
   로컬에서 프로젝트를 테스트하세요.

docs/ko/enterprise/guides/private-package-registry.mdx

@@ -0,0 +1,261 @@
+---
+title: "프라이빗 패키지 레지스트리"
+description: "CrewAI AMP에서 인증된 PyPI 레지스트리의 프라이빗 Python 패키지 설치하기"
+icon: "lock"
+mode: "wide"
+---
+
+<Note>
+  이 가이드는 CrewAI AMP에 배포할 때 프라이빗 PyPI 레지스트리(Azure DevOps Artifacts, GitHub Packages,
+  GitLab, AWS CodeArtifact 등)에서 Python 패키지를 설치하도록 CrewAI 프로젝트를 구성하는 방법을 다룹니다.
+</Note>
+
+## 이 가이드가 필요한 경우
+
+프로젝트가 공개 PyPI가 아닌 프라이빗 레지스트리에 호스팅된 내부 또는 독점 Python 패키지에
+의존하는 경우, 다음을 수행해야 합니다:
+
+1. UV에 패키지를 **어디서** 찾을지 알려줍니다 (index URL)
+2. UV에 **어떤** 패키지가 해당 index에서 오는지 알려줍니다 (source 매핑)
+3. UV가 설치 중에 인증할 수 있도록 **자격 증명**을 제공합니다
+
+CrewAI AMP는 의존성 해결 및 설치에 [UV](https://docs.astral.sh/uv/)를 사용합니다.
+UV는 `pyproject.toml` 구성과 자격 증명용 환경 변수를 결합하여 인증된 프라이빗 레지스트리를 지원합니다.
+
+## 1단계: pyproject.toml 구성
+
+`pyproject.toml`에서 세 가지 요소가 함께 작동합니다:
+
+### 1a. 의존성 선언
+
+프라이빗 패키지를 다른 의존성과 마찬가지로 `[project.dependencies]`에 추가합니다:
+
+```toml
+[project]
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+```
+
+### 1b. index 정의
+
+프라이빗 레지스트리를 `[[tool.uv.index]]` 아래에 명명된 index로 등록합니다:
+
+```toml
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+```
+
+<Info>
+  `name` 필드는 중요합니다 — UV는 이를 사용하여 인증을 위한 환경 변수 이름을
+  구성합니다 (아래 [2단계](#2단계-인증-자격-증명-설정)를 참조하세요).
+
+  `explicit = true`를 설정하면 UV가 모든 패키지에 대해 이 index를 검색하지 않습니다 —
+  `[tool.uv.sources]`에서 명시적으로 매핑한 패키지만 검색합니다. 이렇게 하면 프라이빗
+  레지스트리에 대한 불필요한 쿼리를 방지하고 의존성 혼동 공격을 차단할 수 있습니다.
+</Info>
+
+### 1c. 패키지를 index에 매핑
+
+`[tool.uv.sources]`를 사용하여 프라이빗 index에서 해결해야 할 패키지를 UV에 알려줍니다:
+
+```toml
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+### 전체 예시
+
+```toml
+[project]
+name = "my-crew-project"
+version = "0.1.0"
+requires-python = ">=3.10,<=3.13"
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+
+[tool.crewai]
+type = "crew"
+
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+`pyproject.toml`을 업데이트한 후 lock 파일을 다시 생성합니다:
+
+```bash
+uv lock
+```
+
+<Warning>
+  업데이트된 `uv.lock`을 항상 `pyproject.toml` 변경 사항과 함께 커밋하세요.
+  lock 파일은 배포에 필수입니다 — [배포 준비하기](/ko/enterprise/guides/prepare-for-deployment)를 참조하세요.
+</Warning>
+
+## 2단계: 인증 자격 증명 설정
+
+UV는 `pyproject.toml`에서 정의한 index 이름을 기반으로 한 명명 규칙을 따르는
+환경 변수를 사용하여 프라이빗 index에 인증합니다:
+
+```
+UV_INDEX_{UPPER_NAME}_USERNAME
+UV_INDEX_{UPPER_NAME}_PASSWORD
+```
+
+여기서 `{UPPER_NAME}`은 index 이름을 **대문자**로 변환하고 **하이픈을 언더스코어로 대체**한 것입니다.
+
+예를 들어, `my-private-registry`라는 이름의 index는 다음을 사용합니다:
+
+| 변수 | 값 |
+|------|-----|
+| `UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME` | 레지스트리 사용자 이름 또는 토큰 이름 |
+| `UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD` | 레지스트리 비밀번호 또는 토큰/PAT |
+
+<Warning>
+  이 환경 변수는 CrewAI AMP **환경 변수** 설정을 통해 **반드시** 추가해야 합니다 —
+  전역적으로 또는 배포 수준에서. `.env` 파일에 설정하거나 프로젝트에 하드코딩할 수 없습니다.
+
+  아래 [AMP에서 환경 변수 설정](#amp에서-환경-변수-설정)을 참조하세요.
+</Warning>
+
+## 레지스트리 제공업체 참조
+
+아래 표는 일반적인 레지스트리 제공업체의 index URL 형식과 자격 증명 값을 보여줍니다.
+자리 표시자 값을 실제 조직 및 피드 세부 정보로 대체하세요.
+
+| 제공업체 | Index URL | 사용자 이름 | 비밀번호 |
+|---------|-----------|-----------|---------|
+| **Azure DevOps Artifacts** | `https://pkgs.dev.azure.com/{org}/_packaging/{feed}/pypi/simple/` | 비어 있지 않은 임의의 문자열 (예: `token`) | Packaging Read 범위의 Personal Access Token (PAT) |
+| **GitHub Packages** | `https://pypi.pkg.github.com/{owner}/simple/` | GitHub 사용자 이름 | `read:packages` 범위의 Personal Access Token (classic) |
+| **GitLab Package Registry** | `https://gitlab.com/api/v4/projects/{project_id}/packages/pypi/simple/` | `__token__` | `read_api` 범위의 Project 또는 Personal Access Token |
+| **AWS CodeArtifact** | `aws codeartifact get-repository-endpoint`의 URL 사용 | `aws` | `aws codeartifact get-authorization-token`의 토큰 |
+| **Google Artifact Registry** | `https://{region}-python.pkg.dev/{project}/{repo}/simple/` | `_json_key_base64` | Base64로 인코딩된 서비스 계정 키 |
+| **JFrog Artifactory** | `https://{instance}.jfrog.io/artifactory/api/pypi/{repo}/simple/` | 사용자 이름 또는 이메일 | API 키 또는 ID 토큰 |
+| **자체 호스팅 (devpi, Nexus 등)** | 레지스트리의 simple API URL | 레지스트리 사용자 이름 | 레지스트리 비밀번호 |
+
+<Tip>
+  **AWS CodeArtifact**의 경우 인증 토큰이 주기적으로 만료됩니다.
+  만료되면 `UV_INDEX_*_PASSWORD` 값을 갱신해야 합니다.
+  CI/CD 파이프라인에서 이를 자동화하는 것을 고려하세요.
+</Tip>
+
+## AMP에서 환경 변수 설정
+
+프라이빗 레지스트리 자격 증명은 CrewAI AMP에서 환경 변수로 구성해야 합니다.
+두 가지 옵션이 있습니다:
+
+<Tabs>
+  <Tab title="웹 인터페이스">
+    1. [CrewAI AMP](https://app.crewai.com)에 로그인합니다
+    2. 자동화로 이동합니다
+    3. **Environment Variables** 탭을 엽니다
+    4. 각 변수 (`UV_INDEX_*_USERNAME` 및 `UV_INDEX_*_PASSWORD`)에 값을 추가합니다
+
+    자세한 내용은 [AMP에 배포하기 — 환경 변수 설정하기](/ko/enterprise/guides/deploy-to-amp#환경-변수-설정하기) 단계를 참조하세요.
+  </Tab>
+  <Tab title="CLI 배포">
+    `crewai deploy create`를 실행하기 전에 로컬 `.env` 파일에 변수를 추가합니다.
+    CLI가 이를 안전하게 플랫폼으로 전송합니다:
+
+    ```bash
+    # .env
+    OPENAI_API_KEY=sk-...
+    UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+    UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat-here
+    ```
+
+    ```bash
+    crewai deploy create
+    ```
+  </Tab>
+</Tabs>
+
+<Warning>
+  자격 증명을 저장소에 **절대** 커밋하지 마세요. 모든 비밀 정보에는 AMP 환경 변수를 사용하세요.
+  `.env` 파일은 `.gitignore`에 포함되어야 합니다.
+</Warning>
+
+기존 배포의 자격 증명을 업데이트하려면 [Crew 업데이트하기 — 환경 변수](/ko/enterprise/guides/update-crew)를 참조하세요.
+
+## 전체 동작 흐름
+
+CrewAI AMP가 자동화를 빌드할 때, 해결 흐름은 다음과 같이 작동합니다:
+
+<Steps>
+  <Step title="빌드 시작">
+    AMP가 저장소를 가져오고 `pyproject.toml`과 `uv.lock`을 읽습니다.
+  </Step>
+  <Step title="UV가 의존성 해결">
+    UV가 `[tool.uv.sources]`를 읽어 각 패키지가 어떤 index에서 와야 하는지 결정합니다.
+  </Step>
+  <Step title="UV가 인증">
+    각 프라이빗 index에 대해 UV가 AMP에서 구성한 환경 변수에서
+    `UV_INDEX_{NAME}_USERNAME`과 `UV_INDEX_{NAME}_PASSWORD`를 조회합니다.
+  </Step>
+  <Step title="패키지 설치">
+    UV가 공개(PyPI) 및 프라이빗(레지스트리) 패키지를 모두 다운로드하고 설치합니다.
+  </Step>
+  <Step title="자동화 실행">
+    모든 의존성이 사용 가능한 상태에서 crew 또는 flow가 시작됩니다.
+  </Step>
+</Steps>
+
+## 문제 해결
+
+### 빌드 중 인증 오류
+
+**증상**: 프라이빗 패키지를 해결할 때 `401 Unauthorized` 또는 `403 Forbidden`으로 빌드가 실패합니다.
+
+**확인사항**:
+- `UV_INDEX_*` 환경 변수 이름이 index 이름과 정확히 일치하는지 확인합니다 (대문자, 하이픈 -> 언더스코어)
+- 자격 증명이 로컬 `.env`뿐만 아니라 AMP 환경 변수에 설정되어 있는지 확인합니다
+- 토큰/PAT에 패키지 피드에 필요한 읽기 권한이 있는지 확인합니다
+- 토큰이 만료되지 않았는지 확인합니다 (특히 AWS CodeArtifact의 경우)
+
+### 패키지를 찾을 수 없음
+
+**증상**: `No matching distribution found for my-private-package`.
+
+**확인사항**:
+- `pyproject.toml`의 index URL이 `/simple/`로 끝나는지 확인합니다
+- `[tool.uv.sources]` 항목이 올바른 패키지 이름을 올바른 index 이름에 매핑하는지 확인합니다
+- 패키지가 실제로 프라이빗 레지스트리에 게시되어 있는지 확인합니다
+- 동일한 자격 증명으로 로컬에서 `uv lock`을 실행하여 해결이 작동하는지 확인합니다
+
+### Lock 파일 충돌
+
+**증상**: 프라이빗 index를 추가한 후 `uv lock`이 실패하거나 예상치 못한 결과를 생성합니다.
+
+**해결책**: 로컬에서 자격 증명을 설정하고 다시 생성합니다:
+
+```bash
+export UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+export UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat
+uv lock
+```
+
+그런 다음 업데이트된 `uv.lock`을 커밋합니다.
+
+## 관련 가이드
+
+<CardGroup cols={3}>
+  <Card title="배포 준비하기" icon="clipboard-check" href="/ko/enterprise/guides/prepare-for-deployment">
+    배포 전에 프로젝트 구조와 의존성을 확인합니다.
+  </Card>
+  <Card title="AMP에 배포하기" icon="rocket" href="/ko/enterprise/guides/deploy-to-amp">
+    crew 또는 flow를 배포하고 환경 변수를 구성합니다.
+  </Card>
+  <Card title="Crew 업데이트하기" icon="arrows-rotate" href="/ko/enterprise/guides/update-crew">
+    환경 변수를 업데이트하고 실행 중인 배포에 변경 사항을 푸시합니다.
+  </Card>
+</CardGroup>

docs/ko/guides/migration/migrating-from-langgraph.mdx

@@ -0,0 +1,518 @@
+---
+title: "LangGraph에서 CrewAI로 옮기기: 엔지니어를 위한 실전 가이드"
+description: LangGraph로 이미 구축했다면, 프로젝트를 CrewAI로 빠르게 옮기는 방법을 알아보세요
+icon: switch
+mode: "wide"
+---
+
+LangGraph로 에이전트를 구축해 왔습니다. `StateGraph`와 씨름하고, 조건부 에지를 연결하고, 새벽 2시에 상태 딕셔너리를 디버깅해 본 적도 있죠. 동작은 하지만 — 어느 순간부터 프로덕션으로 가는 더 나은 길이 없을까 고민하게 됩니다.
+
+있습니다. **CrewAI Flows**는 이벤트 기반 오케스트레이션, 조건부 라우팅, 공유 상태라는 동일한 힘을 훨씬 적은 보일러플레이트와 실제로 다단계 AI 워크플로우를 생각하는 방식에 잘 맞는 정신적 모델로 제공합니다.
+
+이 글은 핵심 개념을 나란히 비교하고 실제 코드 비교를 보여주며, 다음으로 손이 갈 프레임워크가 왜 CrewAI Flows인지 설명합니다.
+
+---
+
+## 정신적 모델의 전환
+
+LangGraph는 **그래프**로 생각하라고 요구합니다: 노드, 에지, 그리고 상태 딕셔너리. 모든 워크플로우는 계산 단계 사이의 전이를 명시적으로 연결하는 방향 그래프입니다. 강력하지만, 특히 워크플로우가 몇 개의 결정 지점이 있는 순차적 흐름일 때 이 추상화는 오버헤드를 가져옵니다.
+
+CrewAI Flows는 **이벤트**로 생각하라고 요구합니다: 시작하는 메서드, 결과를 듣는 메서드, 실행을 라우팅하는 메서드. 워크플로우의 토폴로지는 명시적 그래프 구성 대신 데코레이터 어노테이션에서 드러납니다. 이것은 단순한 문법 설탕이 아니라 — 파이프라인을 설계하고 읽고 유지하는 방식을 바꿉니다.
+
+핵심 매핑은 다음과 같습니다:
+
+| LangGraph 개념 | CrewAI Flows 대응 |
+| --- | --- |
+| `StateGraph` class | `Flow` class |
+| `add_node()` | Methods decorated with `@start`, `@listen` |
+| `add_edge()` / `add_conditional_edges()` | `@listen()` / `@router()` decorators |
+| `TypedDict` state | Pydantic `BaseModel` state |
+| `START` / `END` constants | `@start()` decorator / natural method return |
+| `graph.compile()` | `flow.kickoff()` |
+| Checkpointer / persistence | Built-in memory (LanceDB-backed) |
+
+실제로 어떻게 보이는지 살펴보겠습니다.
+
+---
+
+## 데모 1: 간단한 순차 파이프라인
+
+주제를 받아 조사하고, 요약을 작성한 뒤, 결과를 포맷팅하는 파이프라인을 만든다고 해봅시다. 각 프레임워크는 이렇게 처리합니다.
+
+### LangGraph 방식
+
+```python
+from typing import TypedDict
+from langgraph.graph import StateGraph, START, END
+
+class ResearchState(TypedDict):
+    topic: str
+    raw_research: str
+    summary: str
+    formatted_output: str
+
+def research_topic(state: ResearchState) -> dict:
+    # Call an LLM or search API
+    result = llm.invoke(f"Research the topic: {state['topic']}")
+    return {"raw_research": result}
+
+def write_summary(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Summarize this research:\n{state['raw_research']}"
+    )
+    return {"summary": result}
+
+def format_output(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Format this summary as a polished article section:\n{state['summary']}"
+    )
+    return {"formatted_output": result}
+
+# Build the graph
+graph = StateGraph(ResearchState)
+graph.add_node("research", research_topic)
+graph.add_node("summarize", write_summary)
+graph.add_node("format", format_output)
+
+graph.add_edge(START, "research")
+graph.add_edge("research", "summarize")
+graph.add_edge("summarize", "format")
+graph.add_edge("format", END)
+
+# Compile and run
+app = graph.compile()
+result = app.invoke({"topic": "quantum computing advances in 2026"})
+print(result["formatted_output"])
+```
+
+함수를 정의하고 노드로 등록한 다음, 모든 전이를 수동으로 연결합니다. 이렇게 단순한 순서인데도 의례처럼 해야 할 작업이 많습니다.
+
+### CrewAI Flows 방식
+
+```python
+from crewai import LLM, Agent, Crew, Process, Task
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ResearchState(BaseModel):
+    topic: str = ""
+    raw_research: str = ""
+    summary: str = ""
+    formatted_output: str = ""
+
+class ResearchFlow(Flow[ResearchState]):
+    @start()
+    def research_topic(self):
+        # Option 1: Direct LLM call
+        result = llm.call(f"Research the topic: {self.state.topic}")
+        self.state.raw_research = result
+        return result
+
+    @listen(research_topic)
+    def write_summary(self, research_output):
+        # Option 2: A single agent
+        summarizer = Agent(
+            role="Research Summarizer",
+            goal="Produce concise, accurate summaries of research content",
+            backstory="You are an expert at distilling complex research into clear, "
+            "digestible summaries.",
+            llm=llm,
+            verbose=True,
+        )
+        result = summarizer.kickoff(
+            f"Summarize this research:\n{self.state.raw_research}"
+        )
+        self.state.summary = str(result)
+        return self.state.summary
+
+    @listen(write_summary)
+    def format_output(self, summary_output):
+        # Option 3: a complete crew (with one or more agents)
+        formatter = Agent(
+            role="Content Formatter",
+            goal="Transform research summaries into polished, publication-ready article sections",
+            backstory="You are a skilled editor with expertise in structuring and "
+            "presenting technical content for a general audience.",
+            llm=llm,
+            verbose=True,
+        )
+        format_task = Task(
+            description=f"Format this summary as a polished article section:\n{self.state.summary}",
+            expected_output="A well-structured, polished article section ready for publication.",
+            agent=formatter,
+        )
+        crew = Crew(
+            agents=[formatter],
+            tasks=[format_task],
+            process=Process.sequential,
+            verbose=True,
+        )
+        result = crew.kickoff()
+        self.state.formatted_output = str(result)
+        return self.state.formatted_output
+
+# Run the flow
+flow = ResearchFlow()
+flow.state.topic = "quantum computing advances in 2026"
+result = flow.kickoff()
+print(flow.state.formatted_output)
+
+```
+
+눈에 띄는 차이점이 있습니다: 그래프 구성 없음, 에지 연결 없음, 컴파일 단계 없음. 실행 순서는 로직이 있는 곳에서 바로 선언됩니다. `@start()`는 진입점을 표시하고, `@listen(method_name)`은 단계들을 연결합니다. 상태는 타입 안전성, 검증, IDE 자동 완성까지 제공하는 제대로 된 Pydantic 모델입니다.
+
+---
+
+## 데모 2: 조건부 라우팅
+
+여기서 흥미로워집니다. 콘텐츠 유형에 따라 서로 다른 처리 경로로 라우팅하는 파이프라인을 만든다고 해봅시다.
+
+### LangGraph 방식
+
+```python
+from typing import TypedDict, Literal
+from langgraph.graph import StateGraph, START, END
+
+class ContentState(TypedDict):
+    input_text: str
+    content_type: str
+    result: str
+
+def classify_content(state: ContentState) -> dict:
+    content_type = llm.invoke(
+        f"Classify this content as 'technical', 'creative', or 'business':\n{state['input_text']}"
+    )
+    return {"content_type": content_type.strip().lower()}
+
+def process_technical(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as technical doc:\n{state['input_text']}")
+    return {"result": result}
+
+def process_creative(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as creative writing:\n{state['input_text']}")
+    return {"result": result}
+
+def process_business(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as business content:\n{state['input_text']}")
+    return {"result": result}
+
+# Routing function
+def route_content(state: ContentState) -> Literal["technical", "creative", "business"]:
+    return state["content_type"]
+
+# Build the graph
+graph = StateGraph(ContentState)
+graph.add_node("classify", classify_content)
+graph.add_node("technical", process_technical)
+graph.add_node("creative", process_creative)
+graph.add_node("business", process_business)
+
+graph.add_edge(START, "classify")
+graph.add_conditional_edges(
+    "classify",
+    route_content,
+    {
+        "technical": "technical",
+        "creative": "creative",
+        "business": "business",
+    }
+)
+graph.add_edge("technical", END)
+graph.add_edge("creative", END)
+graph.add_edge("business", END)
+
+app = graph.compile()
+result = app.invoke({"input_text": "Explain how TCP handshakes work"})
+```
+
+별도의 라우팅 함수, 명시적 조건부 에지 매핑, 그리고 모든 분기에 대한 종료 에지가 필요합니다. 라우팅 결정 로직이 그 결정을 만들어 내는 노드와 분리됩니다.
+
+### CrewAI Flows 방식
+
+```python
+from crewai import LLM, Agent
+from crewai.flow.flow import Flow, listen, router, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ContentState(BaseModel):
+    input_text: str = ""
+    content_type: str = ""
+    result: str = ""
+
+class ContentFlow(Flow[ContentState]):
+    @start()
+    def classify_content(self):
+        self.state.content_type = (
+            llm.call(
+                f"Classify this content as 'technical', 'creative', or 'business':\n"
+                f"{self.state.input_text}"
+            )
+            .strip()
+            .lower()
+        )
+        return self.state.content_type
+
+    @router(classify_content)
+    def route_content(self, classification):
+        if classification == "technical":
+            return "process_technical"
+        elif classification == "creative":
+            return "process_creative"
+        else:
+            return "process_business"
+
+    @listen("process_technical")
+    def handle_technical(self):
+        agent = Agent(
+            role="Technical Writer",
+            goal="Produce clear, accurate technical documentation",
+            backstory="You are an expert technical writer who specializes in "
+            "explaining complex technical concepts precisely.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as technical doc:\n{self.state.input_text}")
+        )
+
+    @listen("process_creative")
+    def handle_creative(self):
+        agent = Agent(
+            role="Creative Writer",
+            goal="Craft engaging and imaginative creative content",
+            backstory="You are a talented creative writer with a flair for "
+            "compelling storytelling and vivid expression.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as creative writing:\n{self.state.input_text}")
+        )
+
+    @listen("process_business")
+    def handle_business(self):
+        agent = Agent(
+            role="Business Writer",
+            goal="Produce professional, results-oriented business content",
+            backstory="You are an experienced business writer who communicates "
+            "strategy and value clearly to professional audiences.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as business content:\n{self.state.input_text}")
+        )
+
+flow = ContentFlow()
+flow.state.input_text = "Explain how TCP handshakes work"
+flow.kickoff()
+print(flow.state.result)
+
+```
+
+`@router()` 데코레이터는 메서드를 결정 지점으로 만듭니다. 리스너와 매칭되는 문자열을 반환하므로, 매핑 딕셔너리도, 별도의 라우팅 함수도 필요 없습니다. 분기 로직이 Python `if` 문처럼 읽히는 이유는, 실제로 `if` 문이기 때문입니다.
+
+---
+
+## 데모 3: AI 에이전트 Crew를 Flow에 통합하기
+
+여기서 CrewAI의 진짜 힘이 드러납니다. Flows는 LLM 호출을 연결하는 것에 그치지 않고 자율적인 에이전트 **Crew** 전체를 오케스트레이션합니다. 이는 LangGraph에 기본으로 대응되는 개념이 없습니다.
+
+```python
+from crewai import Agent, Task, Crew
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+class ArticleState(BaseModel):
+    topic: str = ""
+    research: str = ""
+    draft: str = ""
+    final_article: str = ""
+
+class ArticleFlow(Flow[ArticleState]):
+
+    @start()
+    def run_research_crew(self):
+        """A full Crew of agents handles research."""
+        researcher = Agent(
+            role="Senior Research Analyst",
+            goal=f"Produce comprehensive research on: {self.state.topic}",
+            backstory="You're a veteran analyst known for thorough, "
+                       "well-sourced research reports.",
+            llm="gpt-4o"
+        )
+
+        research_task = Task(
+            description=f"Research '{self.state.topic}' thoroughly. "
+                        "Cover key trends, data points, and expert opinions.",
+            expected_output="A detailed research brief with sources.",
+            agent=researcher
+        )
+
+        crew = Crew(agents=[researcher], tasks=[research_task])
+        result = crew.kickoff()
+        self.state.research = result.raw
+        return result.raw
+
+    @listen(run_research_crew)
+    def run_writing_crew(self, research_output):
+        """A different Crew handles writing."""
+        writer = Agent(
+            role="Technical Writer",
+            goal="Write a compelling article based on provided research.",
+            backstory="You turn complex research into engaging, clear prose.",
+            llm="gpt-4o"
+        )
+
+        editor = Agent(
+            role="Senior Editor",
+            goal="Review and polish articles for publication quality.",
+            backstory="20 years of editorial experience at top tech publications.",
+            llm="gpt-4o"
+        )
+
+        write_task = Task(
+            description=f"Write an article based on this research:\n{self.state.research}",
+            expected_output="A well-structured draft article.",
+            agent=writer
+        )
+
+        edit_task = Task(
+            description="Review, fact-check, and polish the draft article.",
+            expected_output="A publication-ready article.",
+            agent=editor
+        )
+
+        crew = Crew(agents=[writer, editor], tasks=[write_task, edit_task])
+        result = crew.kickoff()
+        self.state.final_article = result.raw
+        return result.raw
+
+# Run the full pipeline
+flow = ArticleFlow()
+flow.state.topic = "The Future of Edge AI"
+flow.kickoff()
+print(flow.state.final_article)
+```
+
+핵심 인사이트는 다음과 같습니다: **Flows는 오케스트레이션 레이어를, Crews는 지능 레이어를 제공합니다.** Flow의 각 단계는 각자의 역할, 목표, 도구를 가진 협업 에이전트 팀을 띄울 수 있습니다. 구조화되고 예측 가능한 제어 흐름 *그리고* 자율적 에이전트 협업 — 두 세계의 장점을 모두 얻습니다.
+
+LangGraph에서 비슷한 것을 하려면 노드 함수 안에 에이전트 통신 프로토콜, 도구 호출 루프, 위임 로직을 직접 구현해야 합니다. 가능하긴 하지만, 매번 처음부터 배관을 만드는 셈입니다.
+
+---
+
+## 데모 4: 병렬 실행과 동기화
+
+실제 파이프라인은 종종 작업을 병렬로 분기하고 결과를 합쳐야 합니다. CrewAI Flows는 `and_`와 `or_` 연산자로 이를 우아하게 처리합니다.
+
+```python
+from crewai import LLM
+from crewai.flow.flow import Flow, and_, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class AnalysisState(BaseModel):
+    topic: str = ""
+    market_data: str = ""
+    tech_analysis: str = ""
+    competitor_intel: str = ""
+    final_report: str = ""
+
+class ParallelAnalysisFlow(Flow[AnalysisState]):
+    @start()
+    def start_method(self):
+        pass
+
+    @listen(start_method)
+    def gather_market_data(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def run_tech_analysis(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def gather_competitor_intel(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(and_(gather_market_data, run_tech_analysis, gather_competitor_intel))
+    def synthesize_report(self):
+        # Your agentic or deterministic code
+        pass
+
+flow = ParallelAnalysisFlow()
+flow.state.topic = "AI-powered developer tools"
+flow.kickoff()
+
+```
+
+여러 `@start()` 데코레이터는 병렬로 실행됩니다. `@listen` 데코레이터의 `and_()` 결합자는 `synthesize_report`가 *세 가지* 상위 메서드가 모두 완료된 뒤에만 실행되도록 보장합니다. *어떤* 상위 작업이든 끝나는 즉시 진행하고 싶다면 `or_()`도 사용할 수 있습니다.
+
+LangGraph에서는 병렬 분기, 동기화 노드, 신중한 상태 병합이 포함된 fan-out/fan-in 패턴을 만들어야 하며 — 모든 것을 에지로 명시적으로 연결해야 합니다.
+
+---
+
+## 프로덕션에서 CrewAI Flows를 쓰는 이유
+
+깔끔한 문법을 넘어, Flows는 여러 프로덕션 핵심 이점을 제공합니다:
+
+**내장 상태 지속성.** Flow 상태는 LanceDB에 의해 백업되므로 워크플로우가 크래시에서 살아남고, 재개될 수 있으며, 실행 간에 지식을 축적할 수 있습니다. LangGraph는 별도의 체크포인터를 구성해야 합니다.
+
+**타입 안전한 상태 관리.** Pydantic 모델은 즉시 검증, 직렬화, IDE 지원을 제공합니다. LangGraph의 `TypedDict` 상태는 런타임 검증을 하지 않습니다.
+
+**일급 에이전트 오케스트레이션.** Crews는 기본 프리미티브입니다. 역할, 목표, 배경, 도구를 가진 에이전트를 정의하고, Flow의 구조적 틀 안에서 자율적으로 협업하게 합니다. 다중 에이전트 조율을 다시 만들 필요가 없습니다.
+
+**더 단순한 정신적 모델.** 데코레이터는 의도를 선언합니다. `@start`는 "여기서 시작", `@listen(x)`는 "x 이후 실행", `@router(x)`는 "x 이후 어디로 갈지 결정"을 의미합니다. 코드는 자신이 설명하는 워크플로우처럼 읽힙니다.
+
+**CLI 통합.** `crewai run`으로 Flows를 실행합니다. 별도의 컴파일 단계나 그래프 직렬화가 없습니다. Flow는 Python 클래스이며, 그대로 실행됩니다.
+
+---
+
+## 마이그레이션 치트 시트
+
+LangGraph 코드베이스를 CrewAI Flows로 옮기고 싶다면, 다음의 실전 변환 가이드를 참고하세요:
+
+1. **상태를 매핑하세요.** `TypedDict`를 Pydantic `BaseModel`로 변환하고 모든 필드에 기본값을 추가하세요.
+2. **노드를 메서드로 변환하세요.** 각 `add_node` 함수는 `Flow` 서브클래스의 메서드가 됩니다. `state["field"]` 읽기는 `self.state.field`로 바꾸세요.
+3. **에지를 데코레이터로 교체하세요.** `add_edge(START, "first_node")`는 첫 메서드의 `@start()`가 됩니다. 순차적인 `add_edge("a", "b")`는 `b` 메서드의 `@listen(a)`가 됩니다.
+4. **조건부 에지는 `@router`로 교체하세요.** 라우팅 함수와 `add_conditional_edges()` 매핑은 하나의 `@router()` 메서드로 통합하고, 라우트 문자열을 반환하세요.
+5. **compile + invoke를 kickoff으로 교체하세요.** `graph.compile()`를 제거하고 `flow.kickoff()`를 호출하세요.
+6. **Crew가 들어갈 지점을 고려하세요.** 복잡한 다단계 에이전트 로직이 있는 노드는 Crew로 분리할 후보입니다. 이 부분에서 가장 큰 품질 향상을 체감할 수 있습니다.
+
+---
+
+## 시작하기
+
+CrewAI를 설치하고 새 Flow 프로젝트를 스캐폴딩하세요:
+
+```bash
+pip install crewai
+crewai create flow my_first_flow
+cd my_first_flow
+```
+
+이렇게 하면 바로 편집 가능한 Flow 클래스, 설정 파일, 그리고 `type = "flow"`가 이미 설정된 `pyproject.toml`이 포함된 프로젝트 구조가 생성됩니다. 다음으로 실행하세요:
+
+```bash
+crewai run
+```
+
+그 다음부터는 에이전트를 추가하고 리스너를 연결한 뒤, 배포하면 됩니다.
+
+---
+
+## 마무리
+
+LangGraph는 AI 워크플로우에 구조가 필요하다는 사실을 생태계에 일깨워 주었습니다. 중요한 교훈이었습니다. 하지만 CrewAI Flows는 그 교훈을 더 빠르게 쓰고, 더 쉽게 읽으며, 프로덕션에서 더 강력한 형태로 제공합니다 — 특히 워크플로우에 여러 에이전트의 협업이 포함될 때 그렇습니다.
+
+단일 에이전트 체인을 넘는 무엇인가를 만들고 있다면, Flows를 진지하게 검토해 보세요. 데코레이터 기반 모델, Crews의 네이티브 통합, 내장 상태 관리를 통해 배관 작업에 쓰는 시간을 줄이고, 중요한 문제에 더 많은 시간을 쓸 수 있습니다.
+
+`crewai create flow`로 시작하세요. 후회하지 않을 겁니다.

docs/ko/learn/human-feedback-in-flows.mdx

@@ -98,33 +98,43 @@ def handle_feedback(self, result):
 `emit`을 지정하면, 데코레이터는 라우터가 됩니다. 인간의 자유 형식 피드백이 LLM에 의해 해석되어 지정된 outcome 중 하나로 매핑됩니다:
 
 ```python Code
-@start()
-@human_feedback(
-    message="이 콘텐츠의 출판을 승인하시겠습니까?",
-    emit=["approved", "rejected", "needs_revision"],
-    llm="gpt-4o-mini",
-    default_outcome="needs_revision",
-)
-def review_content(self):
-    return "블로그 게시물 초안 내용..."
+from crewai.flow.flow import Flow, start, listen, or_
+from crewai.flow.human_feedback import human_feedback
 
-@listen("approved")
-def publish(self, result):
-    print(f"출판 중! 사용자 의견: {result.feedback}")
+class ReviewFlow(Flow):
+    @start()
+    def generate_content(self):
+        return "블로그 게시물 초안 내용..."
 
-@listen("rejected")
-def discard(self, result):
-    print(f"폐기됨. 이유: {result.feedback}")
+    @human_feedback(
+        message="이 콘텐츠의 출판을 승인하시겠습니까?",
+        emit=["approved", "rejected", "needs_revision"],
+        llm="gpt-4o-mini",
+        default_outcome="needs_revision",
+    )
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "블로그 게시물 초안 내용..."
 
-@listen("needs_revision")
-def revise(self, result):
-    print(f"다음을 기반으로 수정 중: {result.feedback}")
+    @listen("approved")
+    def publish(self, result):
+        print(f"출판 중! 사용자 의견: {result.feedback}")
+
+    @listen("rejected")
+    def discard(self, result):
+        print(f"폐기됨. 이유: {result.feedback}")
 ```
 
+사용자가 "더 자세한 내용이 필요합니다"와 같이 말하면, LLM이 이를 `"needs_revision"`으로 매핑하고, `or_()`를 통해 `review_content`가 다시 트리거됩니다 — 수정 루프가 생성됩니다. outcome이 `"approved"` 또는 `"rejected"`가 될 때까지 루프가 계속됩니다.
+
 <Tip>
 LLM은 가능한 경우 구조화된 출력(function calling)을 사용하여 응답이 지정된 outcome 중 하나임을 보장합니다. 이로 인해 라우팅이 신뢰할 수 있고 예측 가능해집니다.
 </Tip>
 
+<Warning>
+`@start()` 메서드는 flow 시작 시 한 번만 실행됩니다. 수정 루프가 필요한 경우, start 메서드를 review 메서드와 분리하고 review 메서드에 `@listen(or_("trigger", "revision_outcome"))`를 사용하여 self-loop을 활성화하세요.
+</Warning>
+
 ## HumanFeedbackResult
 
 `HumanFeedbackResult` 데이터클래스는 인간 피드백 상호작용에 대한 모든 정보를 포함합니다:
@@ -193,116 +203,162 @@ def summarize(self):
 <CodeGroup>
 
 ```python Code
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 from pydantic import BaseModel
 
 
 class ContentState(BaseModel):
-    topic: str = ""
     draft: str = ""
-    final_content: str = ""
     revision_count: int = 0
+    status: str = "pending"
 
 
 class ContentApprovalFlow(Flow[ContentState]):
-    """콘텐츠를 생성하고 인간의 승인을 받는 Flow입니다."""
+    """콘텐츠를 생성하고 승인될 때까지 반복하는 Flow."""
 
     @start()
-    def get_topic(self):
-        self.state.topic = input("어떤 주제에 대해 글을 쓸까요? ")
-        return self.state.topic
-
-    @listen(get_topic)
-    def generate_draft(self, topic):
-        # 실제 사용에서는 LLM을 호출합니다
-        self.state.draft = f"# {topic}\n\n{topic}에 대한 초안입니다..."
+    def generate_draft(self):
+        self.state.draft = "# AI 안전\n\nAI 안전에 대한 초안..."
         return self.state.draft
 
-    @listen(generate_draft)
     @human_feedback(
-        message="이 초안을 검토해 주세요. 'approved', 'rejected'로 답하거나 수정 피드백을 제공해 주세요:",
+        message="이 초안을 검토해 주세요. 승인, 거부 또는 변경이 필요한 사항을 설명해 주세요:",
         emit=["approved", "rejected", "needs_revision"],
         llm="gpt-4o-mini",
         default_outcome="needs_revision",
     )
-    def review_draft(self, draft):
-        return draft
+    @listen(or_("generate_draft", "needs_revision"))
+    def review_draft(self):
+        self.state.revision_count += 1
+        return f"{self.state.draft} (v{self.state.revision_count})"
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
-        self.state.final_content = result.output
-        print("\n✅ 콘텐츠가 승인되어 출판되었습니다!")
-        print(f"검토자 코멘트: {result.feedback}")
+        self.state.status = "published"
+        print(f"콘텐츠 승인 및 게시! 리뷰어 의견: {result.feedback}")
         return "published"
 
     @listen("rejected")
     def handle_rejection(self, result: HumanFeedbackResult):
-        print("\n❌ 콘텐츠가 거부되었습니다")
-        print(f"이유: {result.feedback}")
+        self.state.status = "rejected"
+        print(f"콘텐츠 거부됨. 이유: {result.feedback}")
         return "rejected"
 
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        self.state.revision_count += 1
-        print(f"\n📝 수정 #{self.state.revision_count} 요청됨")
-        print(f"피드백: {result.feedback}")
 
-        # 실제 Flow에서는 generate_draft로 돌아갈 수 있습니다
-        # 이 예제에서는 단순히 확인합니다
-        return "revision_requested"
-
-
-# Flow 실행
 flow = ContentApprovalFlow()
 result = flow.kickoff()
-print(f"\nFlow 완료. 요청된 수정: {flow.state.revision_count}")
+print(f"\nFlow 완료. 상태: {flow.state.status}, 검토 횟수: {flow.state.revision_count}")
 ```
 
 ```text Output
-어떤 주제에 대해 글을 쓸까요? AI 안전
+==================================================
+OUTPUT FOR REVIEW:
+==================================================
+# AI 안전
+
+AI 안전에 대한 초안... (v1)
+==================================================
+
+이 초안을 검토해 주세요. 승인, 거부 또는 변경이 필요한 사항을 설명해 주세요:
+(Press Enter to skip, or type your feedback)
+
+Your feedback: 더 자세한 내용이 필요합니다
 
 ==================================================
 OUTPUT FOR REVIEW:
 ==================================================
 # AI 안전
 
-AI 안전에 대한 초안입니다...
+AI 안전에 대한 초안... (v2)
 ==================================================
 
-이 초안을 검토해 주세요. 'approved', 'rejected'로 답하거나 수정 피드백을 제공해 주세요:
+이 초안을 검토해 주세요. 승인, 거부 또는 변경이 필요한 사항을 설명해 주세요:
 (Press Enter to skip, or type your feedback)
 
 Your feedback: 좋아 보입니다, 승인!
 
-✅ 콘텐츠가 승인되어 출판되었습니다!
-검토자 코멘트: 좋아 보입니다, 승인!
+콘텐츠 승인 및 게시! 리뷰어 의견: 좋아 보입니다, 승인!
 
-Flow 완료. 요청된 수정: 0
+Flow 완료. 상태: published, 검토 횟수: 2
 ```
 
 </CodeGroup>
 
 ## 다른 데코레이터와 결합하기
 
-`@human_feedback` 데코레이터는 다른 Flow 데코레이터와 함께 작동합니다. 가장 안쪽 데코레이터(함수에 가장 가까운)로 배치하세요:
+`@human_feedback` 데코레이터는 `@start()`, `@listen()`, `or_()`와 함께 작동합니다. 데코레이터 순서는 두 가지 모두 동작합니다—프레임워크가 양방향으로 속성을 전파합니다—하지만 권장 패턴은 다음과 같습니다:
 
 ```python Code
-# 올바름: @human_feedback이 가장 안쪽(함수에 가장 가까움)
+# Flow 시작 시 일회성 검토 (self-loop 없음)
 @start()
-@human_feedback(message="이것을 검토해 주세요:")
+@human_feedback(message="이것을 검토해 주세요:", emit=["approved", "rejected"], llm="gpt-4o-mini")
 def my_start_method(self):
     return "content"
 
+# 리스너에서 선형 검토 (self-loop 없음)
 @listen(other_method)
-@human_feedback(message="이것도 검토해 주세요:")
+@human_feedback(message="이것도 검토해 주세요:", emit=["good", "bad"], llm="gpt-4o-mini")
 def my_listener(self, data):
     return f"processed: {data}"
+
+# Self-loop: 수정을 위해 반복할 수 있는 검토
+@human_feedback(message="승인 또는 수정 요청?", emit=["approved", "revise"], llm="gpt-4o-mini")
+@listen(or_("upstream_method", "revise"))
+def review_with_loop(self):
+    return "content for review"
 ```
 
-<Tip>
-`@human_feedback`를 가장 안쪽 데코레이터(마지막/함수에 가장 가까움)로 배치하여 메서드를 직접 래핑하고 Flow 시스템에 전달하기 전에 반환 값을 캡처할 수 있도록 하세요.
-</Tip>
+### Self-loop 패턴
+
+수정 루프를 만들려면 `or_()`를 사용하여 검토 메서드가 **상위 트리거**와 **자체 수정 outcome**을 모두 리스닝해야 합니다:
+
+```python Code
+@start()
+def generate(self):
+    return "initial draft"
+
+@human_feedback(
+    message="승인하시겠습니까, 아니면 변경을 요청하시겠습니까?",
+    emit=["revise", "approved"],
+    llm="gpt-4o-mini",
+    default_outcome="approved",
+)
+@listen(or_("generate", "revise"))
+def review(self):
+    return "content"
+
+@listen("approved")
+def publish(self):
+    return "published"
+```
+
+outcome이 `"revise"`이면 flow가 `review`로 다시 라우팅됩니다 (`or_()`를 통해 `"revise"`를 리스닝하기 때문). outcome이 `"approved"`이면 flow가 `publish`로 계속됩니다. flow 엔진이 라우터를 "한 번만 실행" 규칙에서 제외하여 각 루프 반복마다 재실행할 수 있기 때문에 이 패턴이 동작합니다.
+
+### 체인된 라우터
+
+한 라우터의 outcome으로 트리거된 리스너가 그 자체로 라우터가 될 수 있습니다:
+
+```python Code
+@start()
+@human_feedback(message="첫 번째 검토:", emit=["approved", "rejected"], llm="gpt-4o-mini")
+def draft(self):
+    return "draft content"
+
+@listen("approved")
+@human_feedback(message="최종 검토:", emit=["publish", "revise"], llm="gpt-4o-mini")
+def final_review(self, prev):
+    return "final content"
+
+@listen("publish")
+def on_publish(self, prev):
+    return "published"
+```
+
+### 제한 사항
+
+- **`@start()` 메서드는 한 번만 실행**: `@start()` 메서드는 self-loop할 수 없습니다. 수정 주기가 필요하면 별도의 `@start()` 메서드를 진입점으로 사용하고 `@listen()` 메서드에 `@human_feedback`를 배치하세요.
+- **동일 메서드에 `@start()` + `@listen()` 불가**: 이는 Flow 프레임워크 제약입니다. 메서드는 시작점이거나 리스너여야 하며, 둘 다일 수 없습니다.
 
 ## 모범 사례
 
@@ -516,9 +572,9 @@ class ContentPipeline(Flow):
     @start()
     @human_feedback(
         message="이 콘텐츠의 출판을 승인하시겠습니까?",
-        emit=["approved", "rejected", "needs_revision"],
+        emit=["approved", "rejected"],
         llm="gpt-4o-mini",
-        default_outcome="needs_revision",
+        default_outcome="rejected",
         provider=SlackNotificationProvider("#content-reviews"),
     )
     def generate_content(self):
@@ -534,11 +590,6 @@ class ContentPipeline(Flow):
         print(f"보관됨. 이유: {result.feedback}")
         return {"status": "archived"}
 
-    @listen("needs_revision")
-    def queue_revision(self, result):
-        print(f"수정 대기열에 추가됨: {result.feedback}")
-        return {"status": "revision_needed"}
-
 
 # Flow 시작 (Slack 응답을 기다리며 일시 중지)
 def start_content_pipeline():
@@ -594,22 +645,22 @@ async def on_slack_feedback_async(flow_id: str, slack_message: str):
 ```python Code
 class ArticleReviewFlow(Flow):
     @start()
-    @human_feedback(
-        message="Review this article draft:",
-        emit=["approved", "needs_revision"],
-        llm="gpt-4o-mini",
-        learn=True,  # HITL 학습 활성화
-    )
     def generate_article(self):
         return self.crew.kickoff(inputs={"topic": "AI Safety"}).raw
 
+    @human_feedback(
+        message="이 글 초안을 검토해 주세요:",
+        emit=["approved", "needs_revision"],
+        llm="gpt-4o-mini",
+        learn=True,
+    )
+    @listen(or_("generate_article", "needs_revision"))
+    def review_article(self):
+        return self.last_human_feedback.output if self.last_human_feedback else "article draft"
+
     @listen("approved")
     def publish(self):
         print(f"Publishing: {self.last_human_feedback.output}")
-
-    @listen("needs_revision")
-    def revise(self):
-        print("Revising based on feedback...")
 ```
 
 **첫 번째 실행**: 인간이 원시 출력을 보고 "사실에 대한 주장에는 항상 인용을 포함하세요."라고 말합니다. 교훈이 추출되어 메모리에 저장됩니다.

docs/ko/learn/llm-connections.mdx

@@ -7,7 +7,7 @@ mode: "wide"
 
 ## CrewAI를 LLM에 연결하기
 
-CrewAI는 LiteLLM을 사용하여 다양한 언어 모델(LLM)에 연결합니다. 이 통합은 높은 다양성을 제공하여, 여러 공급자의 모델을 간단하고 통합된 인터페이스로 사용할 수 있게 해줍니다.
+CrewAI는 가장 인기 있는 제공자(OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock)에 대해 네이티브 SDK 통합을 통해 LLM에 연결하며, 그 외 모든 제공자에 대해서는 LiteLLM을 유연한 폴백으로 사용합니다.
 
 <Note>
     기본적으로 CrewAI는 `gpt-4o-mini` 모델을 사용합니다. 이는 `OPENAI_MODEL_NAME` 환경 변수에 의해 결정되며, 설정되지 않은 경우 기본값은 "gpt-4o-mini"입니다.
@@ -41,6 +41,14 @@ LiteLLM은 다음을 포함하되 이에 국한되지 않는 다양한 프로바
 
 지원되는 프로바이더의 전체 및 최신 목록은 [LiteLLM 프로바이더 문서](https://docs.litellm.ai/docs/providers)를 참조하세요.
 
+<Info>
+  네이티브 통합에서 지원하지 않는 제공자를 사용하려면 LiteLLM을 프로젝트에 의존성으로 추가하세요:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+  네이티브 제공자(OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock)는 자체 SDK extras를 사용합니다 — [공급자 구성 예시](/ko/concepts/llms#공급자-구성-예시)를 참조하세요.
+</Info>
+
 ## LLM 변경하기
 
 CrewAI agent에서 다른 LLM을 사용하려면 여러 가지 방법이 있습니다:

docs/ko/observability/tracing.mdx

@@ -35,7 +35,7 @@ crewai login
 아직 설치하지 않았다면 CLI 도구와 함께 CrewAI를 설치하세요:
 
 ```bash
-uv add crewai[tools]
+uv add 'crewai[tools]'
 ```
 
 그런 다음 CrewAI AMP 계정으로 CLI를 인증하세요:

docs/ko/tools/automation/composiotool.mdx

@@ -18,77 +18,46 @@ Composio는 AI 에이전트를 250개 이상의 도구와 연결할 수 있는
 Composio 도구를 프로젝트에 통합하려면 아래 지침을 따르세요:
 
 ```shell
-pip install composio-crewai
+pip install composio composio-crewai
 pip install crewai
 ```
 
-설치가 완료된 후, `composio login`을 실행하거나 Composio API 키를 `COMPOSIO_API_KEY`로 export하세요. Composio API 키는 [여기](https://app.composio.dev)에서 받을 수 있습니다.
+설치가 완료되면 Composio API 키를 `COMPOSIO_API_KEY`로 설정하세요. Composio API 키는 [여기](https://platform.composio.dev)에서 받을 수 있습니다.
 
 ## 예시
 
-다음 예시는 도구를 초기화하고 github action을 실행하는 방법을 보여줍니다:
+다음 예시는 도구를 초기화하고 GitHub 액션을 실행하는 방법을 보여줍니다:
 
-1. Composio 도구 세트 초기화
+1. CrewAI Provider와 함께 Composio 초기화
 
 ```python Code
-from composio_crewai import ComposioToolSet, App, Action
+from composio_crewai import ComposioProvider
+from composio import Composio
 from crewai import Agent, Task, Crew
 
-toolset = ComposioToolSet()
+composio = Composio(provider=ComposioProvider())
 ```
 
-2. GitHub 계정 연결
+2. 새 Composio 세션을 만들고 도구 가져오기
 <CodeGroup>
-```shell CLI
-composio add github
-```
-```python Code
-request = toolset.initiate_connection(app=App.GITHUB)
-print(f"Open this URL to authenticate: {request.redirectUrl}")
+```python
+session = composio.create(
+    user_id="your-user-id",
+    toolkits=["gmail", "github"] # optional, default is all toolkits
+)
+tools = session.tools()
 ```
+세션 및 사용자 관리에 대한 자세한 내용은 [여기](https://docs.composio.dev/docs/configuring-sessions)를 참고하세요.
 </CodeGroup>
 
-3. 도구 가져오기
+3. 사용자 수동 인증하기
 
-- 앱에서 모든 도구를 가져오기 (프로덕션 환경에서는 권장하지 않음):
+Composio는 에이전트 채팅 세션 중에 사용자를 자동으로 인증합니다. 하지만 `authorize` 메서드를 호출해 사용자를 수동으로 인증할 수도 있습니다.
 ```python Code
-tools = toolset.get_tools(apps=[App.GITHUB])
+connection_request = session.authorize("github")
+print(f"Open this URL to authenticate: {connection_request.redirect_url}")
 ```
 
-- 태그를 기반으로 도구 필터링:
-```python Code
-tag = "users"
-
-filtered_action_enums = toolset.find_actions_by_tags(
-    App.GITHUB,
-    tags=[tag], 
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-
-- 사용 사례를 기반으로 도구 필터링:
-```python Code
-use_case = "Star a repository on GitHub"
-
-filtered_action_enums = toolset.find_actions_by_use_case(
-    App.GITHUB, use_case=use_case, advanced=False
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-<Tip>`advanced`를 True로 설정하면 복잡한 사용 사례를 위한 액션을 가져올 수 있습니다</Tip>
-
-- 특정 도구 사용하기:
-
-이 데모에서는 GitHub 앱의 `GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER` 액션을 사용합니다.
-```python Code
-tools = toolset.get_tools(
-    actions=[Action.GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER]
-)
-```
-액션 필터링에 대해 더 자세한 내용을 보려면 [여기](https://docs.composio.dev/patterns/tools/use-tools/use-specific-actions)를 참고하세요.
-
 4. 에이전트 정의
 
 ```python Code
@@ -116,4 +85,4 @@ crew = Crew(agents=[crewai_agent], tasks=[task])
 crew.kickoff()
 ```
 
-* 더욱 자세한 도구 리스트는 [여기](https://app.composio.dev)에서 확인하실 수 있습니다.
+* 더욱 자세한 도구 목록은 [여기](https://docs.composio.dev/toolkits)에서 확인할 수 있습니다.

docs/pt-BR/changelog.mdx

@@ -4,6 +4,138 @@ description: "Atualizações de produto, melhorias e correções do CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="04 mar 2026">
+  ## v1.10.1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## O que mudou
+
+  ### Recursos
+  - Atualizar Gemini GenAI
+
+  ### Correções de Bugs
+  - Ajustar o valor do listener do executor para evitar recursão
+  - Agrupar partes da resposta da função paralela em um único objeto Content no Gemini
+  - Exibir a saída de pensamento dos modelos de pensamento no Gemini
+  - Carregar ferramentas MCP e da plataforma quando as ferramentas do agente forem None
+  - Suportar ambientes Jupyter com loops de eventos em A2A
+  - Usar ID anônimo para rastreamentos efêmeros
+  - Passar condicionalmente o cabeçalho plus
+  - Ignorar o registro do manipulador de sinal em threads não principais para telemetria
+  - Injetar erros de ferramentas como observações e resolver colisões de nomes
+  - Atualizar pypdf de 4.x para 6.7.4 para resolver alertas do Dependabot
+  - Resolver alertas de segurança críticos e altos do Dependabot
+
+  ### Documentação
+  - Sincronizar a documentação da ferramenta Composio entre locais
+
+  ## Contribuidores
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
+<Update label="27 fev 2026">
+  ## v1.10.1a1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## O que Mudou
+
+  ### Funcionalidades
+  - Implementar suporte a invocação assíncrona em métodos de callback de etapas
+  - Implementar carregamento sob demanda para dependências pesadas no módulo de Memória
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.10.0
+
+  ### Refatoração
+  - Refatorar métodos de callback de etapas para suportar invocação assíncrona
+  - Refatorar para implementar carregamento sob demanda para dependências pesadas no módulo de Memória
+
+  ### Correções de Bugs
+  - Corrigir branch para notas de lançamento
+
+  ## Contribuidores
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="27 fev 2026">
+  ## v1.10.1a1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1a1)
+
+  ## O que Mudou
+
+  ### Refatoração
+  - Refatorar métodos de callback de etapas para suportar invocação assíncrona
+  - Implementar carregamento sob demanda para dependências pesadas no módulo de Memória
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.10.0
+
+  ### Correções de Bugs
+  - Criar branch para notas de lançamento
+
+  ## Contribuidores
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="26 fev 2026">
+  ## v1.10.0
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.0)
+
+  ## O que Mudou
+
+  ### Recursos
+  - Aprimorar a resolução da ferramenta MCP e eventos relacionados
+  - Atualizar a versão do lancedb e adicionar pacotes lance-namespace
+  - Aprimorar a análise e validação de argumentos JSON no CrewAgentExecutor e BaseTool
+  - Migrar o cliente HTTP da CLI de requests para httpx
+  - Adicionar documentação versionada
+  - Adicionar detecção de versões removidas para notas de versão
+  - Implementar tratamento de entrada do usuário em Flows
+  - Aprimorar a funcionalidade de auto-loop HITL nos testes de integração de feedback humano
+  - Adicionar started_event_id e definir no eventbus
+  - Atualizar automaticamente tools.specs
+
+  ### Correções de Bugs
+  - Validar kwargs da ferramenta mesmo quando vazios para evitar TypeError crípticos
+  - Preservar tipos nulos nos esquemas de parâmetros da ferramenta para LLM
+  - Mapear output_pydantic/output_json para saída estruturada nativa
+  - Garantir que callbacks sejam executados/aguardados se forem promessas
+  - Capturar o nome do método no contexto da exceção
+  - Preservar tipo enum no resultado do roteador; melhorar tipos
+  - Corrigir fluxos cíclicos que quebram silenciosamente quando o ID de persistência é passado nas entradas
+  - Corrigir o formato da flag da CLI de --skip-provider para --skip_provider
+  - Garantir que o fluxo de chamada da ferramenta OpenAI seja finalizado
+  - Resolver ponteiros $ref de esquema complexos nas ferramentas MCP
+  - Impor additionalProperties=false nos esquemas
+  - Rejeitar nomes de scripts reservados para pastas de equipe
+  - Resolver condição de corrida no teste de emissão de eventos de guardrail
+
+  ### Documentação
+  - Adicionar nota de dependência litellm para provedores de LLM não nativos
+  - Esclarecer o modelo de segurança NL2SQL e orientações de fortalecimento
+  - Adicionar 96 ações ausentes em 9 integrações
+
+  ### Refatoração
+  - Refatorar crew para provider
+  - Extrair HITL para padrão de provider
+  - Melhorar tipagem e registro de hooks
+
+  ## Contribuidores
+
+  @dependabot[bot], @github-actions[bot], @github-code-quality[bot], @greysonlalonde, @heitorado, @hobostay, @joaomdmoura, @johnvan7, @jonathansampson, @lorenzejay, @lucasgomide, @mattatcha, @mplachta, @nicoferdi96, @theCyberTech, @thiagomoretto, @vinibrsl
+
+</Update>
+
 <Update label="26 jan 2026">
   ## v1.9.0
 

docs/pt-BR/concepts/llms.mdx

@@ -105,6 +105,15 @@ Existem diferentes locais no código do CrewAI onde você pode especificar o mod
   </Tab>
 </Tabs>
 
+<Info>
+  O CrewAI oferece integrações nativas via SDK para OpenAI, Anthropic, Google (Gemini API), Azure e AWS Bedrock — sem necessidade de instalação extra além dos extras específicos do provedor (ex.: `uv add "crewai[openai]"`).
+
+  Todos os outros provedores são alimentados pelo **LiteLLM**. Se você planeja usar algum deles, adicione-o como dependência ao seu projeto:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+</Info>
+
 ## Exemplos de Configuração de Provedores
 
 O CrewAI suporta uma grande variedade de provedores de LLM, cada um com recursos, métodos de autenticação e capacidades de modelo únicos.
@@ -214,6 +223,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
     | `meta_llama/Llama-4-Maverick-17B-128E-Instruct-FP8` | 128k | 4028 | Texto, Imagem | Texto |
     | `meta_llama/Llama-3.3-70B-Instruct` | 128k | 4028 | Texto | Texto |
     | `meta_llama/Llama-3.3-8B-Instruct` | 128k | 4028 | Texto | Texto |
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Anthropic">
@@ -354,6 +368,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
     | gemini-1.5-flash                 | 1M tokens          | Modelo multimodal equilibrado, bom para maioria das tarefas         |
     | gemini-1.5-flash-8B              | 1M tokens          | Mais rápido, mais eficiente em custo, adequado para tarefas de alta frequência |
     | gemini-1.5-pro                   | 2M tokens          | Melhor desempenho para uma ampla variedade de tarefas de raciocínio, incluindo lógica, codificação e colaboração criativa |
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Azure">
@@ -438,6 +457,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         model="sagemaker/<my-endpoint>"
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Mistral">
@@ -453,6 +477,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         temperature=0.7
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Nvidia NIM">
@@ -539,6 +568,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
     | rakuten/rakutenai-7b-instruct                                            | 1.024 tokens       | LLM topo de linha, compreensão, raciocínio e geração textual.|
     | rakuten/rakutenai-7b-chat                                                | 1.024 tokens       | LLM topo de linha, compreensão, raciocínio e geração textual.|
     | baichuan-inc/baichuan2-13b-chat                                          | 4.096 tokens       | Suporte a chat em chinês/inglês, programação, matemática, seguir instruções, resolver quizzes.|
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Local NVIDIA NIM Deployed using WSL2">
@@ -579,6 +613,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
 
         # ...
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Groq">
@@ -600,6 +639,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
     | Llama 3.1 70B/8B  | 131.072 tokens      | Alta performance e tarefas de contexto grande|
     | Llama 3.2 Série   | 8.192 tokens        | Tarefas gerais                          |
     | Mixtral 8x7B      | 32.768 tokens       | Equilíbrio entre performance e contexto  |
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="IBM watsonx.ai">
@@ -622,6 +666,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         base_url="https://api.watsonx.ai/v1"
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Ollama (LLMs Locais)">
@@ -635,6 +684,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         base_url="http://localhost:11434"
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Fireworks AI">
@@ -650,6 +704,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         temperature=0.7
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Perplexity AI">
@@ -665,6 +724,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         base_url="https://api.perplexity.ai/"
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Hugging Face">
@@ -679,6 +743,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
         model="huggingface/meta-llama/Meta-Llama-3.1-8B-Instruct"
     )
     ```
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="SambaNova">
@@ -702,6 +771,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
     | Llama 3.2 Série   | 8.192 tokens              | Tarefas gerais e multimodais                 |
     | Llama 3.3 70B     | Até 131.072 tokens        | Desempenho e qualidade de saída elevada      |
     | Família Qwen2     | 8.192 tokens              | Desempenho e qualidade de saída elevada      |
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Cerebras">
@@ -727,6 +801,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
       - Equilíbrio entre velocidade e qualidade
       - Suporte a longas janelas de contexto
     </Info>
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 
   <Accordion title="Open Router">
@@ -749,6 +828,11 @@ Nesta seção, você encontrará exemplos detalhados que ajudam a selecionar, co
       - openrouter/deepseek/deepseek-r1
       - openrouter/deepseek/deepseek-chat
     </Info>
+
+    **Nota:** Este provedor usa o LiteLLM. Adicione-o como dependência ao seu projeto:
+    ```bash
+    uv add 'crewai[litellm]'
+    ```
   </Accordion>
 </AccordionGroup>
 

docs/pt-BR/enterprise/features/flow-hitl-management.mdx

@@ -38,22 +38,21 @@ O CrewAI Enterprise oferece um sistema abrangente de gerenciamento Human-in-the-
 Configure checkpoints de revisão humana em seus Flows usando o decorador `@human_feedback`. Quando a execução atinge um ponto de revisão, o sistema pausa, notifica o responsável via email e aguarda uma resposta.
 
 ```python
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 
 class ContentApprovalFlow(Flow):
     @start()
     def generate_content(self):
-        # IA gera conteúdo
         return "Texto de marketing gerado para campanha Q1..."
 
-    @listen(generate_content)
     @human_feedback(
         message="Por favor, revise este conteúdo para conformidade com a marca:",
         emit=["approved", "rejected", "needs_revision"],
     )
-    def review_content(self, content):
-        return content
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "Texto de marketing para revisão..."
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
@@ -62,10 +61,6 @@ class ContentApprovalFlow(Flow):
     @listen("rejected")
     def archive_content(self, result: HumanFeedbackResult):
         print(f"Conteúdo rejeitado. Motivo: {result.feedback}")
-
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        print(f"Revisão solicitada: {result.feedback}")
 ```
 
 Para detalhes completos de implementação, consulte o guia [Feedback Humano em Flows](/pt-BR/learn/human-feedback-in-flows).

docs/pt-BR/enterprise/guides/deploy-to-amp.mdx

@@ -176,6 +176,11 @@ Você precisa enviar seu crew para um repositório do GitHub. Caso ainda não te
       ![Definir Variáveis de Ambiente](/images/enterprise/set-env-variables.png)
     </Frame>
 
+    <Info>
+      Usando pacotes Python privados? Você também precisará adicionar suas credenciais de registro aqui.
+      Consulte [Registros de Pacotes Privados](/pt-BR/enterprise/guides/private-package-registry) para as variáveis necessárias.
+    </Info>
+
   </Step>
 
   <Step title="Implante Seu Crew">

docs/pt-BR/enterprise/guides/prepare-for-deployment.mdx

@@ -256,6 +256,12 @@ Antes da implantação, certifique-se de ter:
 1. **Chaves de API de LLM** prontas (OpenAI, Anthropic, Google, etc.)
 2. **Chaves de API de ferramentas** se estiver usando ferramentas externas (Serper, etc.)
 
+<Info>
+  Se seu projeto depende de pacotes de um **registro PyPI privado**, você também precisará configurar
+  credenciais de autenticação do registro como variáveis de ambiente. Consulte o guia
+  [Registros de Pacotes Privados](/pt-BR/enterprise/guides/private-package-registry) para mais detalhes.
+</Info>
+
 <Tip>
   Teste seu projeto localmente com as mesmas variáveis de ambiente antes de implantar
   para detectar problemas de configuração antecipadamente.

docs/pt-BR/enterprise/guides/private-package-registry.mdx

@@ -0,0 +1,263 @@
+---
+title: "Registros de Pacotes Privados"
+description: "Instale pacotes Python privados de registros PyPI autenticados no CrewAI AMP"
+icon: "lock"
+mode: "wide"
+---
+
+<Note>
+  Este guia aborda como configurar seu projeto CrewAI para instalar pacotes Python
+  de registros PyPI privados (Azure DevOps Artifacts, GitHub Packages, GitLab, AWS CodeArtifact, etc.)
+  ao implantar no CrewAI AMP.
+</Note>
+
+## Quando Você Precisa Disso
+
+Se seu projeto depende de pacotes Python internos ou proprietários hospedados em um registro privado
+em vez do PyPI público, você precisará:
+
+1. Informar ao UV **onde** encontrar o pacote (uma URL de index)
+2. Informar ao UV **quais** pacotes vêm desse index (um mapeamento de source)
+3. Fornecer **credenciais** para que o UV possa autenticar durante a instalação
+
+O CrewAI AMP usa [UV](https://docs.astral.sh/uv/) para resolução e instalação de dependências.
+O UV suporta registros privados autenticados por meio da configuração do `pyproject.toml` combinada
+com variáveis de ambiente para credenciais.
+
+## Passo 1: Configurar o pyproject.toml
+
+Três elementos trabalham juntos no seu `pyproject.toml`:
+
+### 1a. Declarar a dependência
+
+Adicione o pacote privado ao seu `[project.dependencies]` como qualquer outra dependência:
+
+```toml
+[project]
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+```
+
+### 1b. Definir o index
+
+Registre seu registro privado como um index nomeado em `[[tool.uv.index]]`:
+
+```toml
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+```
+
+<Info>
+  O campo `name` é importante — o UV o utiliza para construir os nomes das variáveis de ambiente
+  para autenticação (veja o [Passo 2](#passo-2-configurar-credenciais-de-autenticação) abaixo).
+
+  Definir `explicit = true` significa que o UV não consultará esse index para todos os pacotes — apenas
+  os que você mapear explicitamente em `[tool.uv.sources]`. Isso evita consultas desnecessárias
+  ao seu registro privado e protege contra ataques de confusão de dependências.
+</Info>
+
+### 1c. Mapear o pacote para o index
+
+Informe ao UV quais pacotes devem ser resolvidos a partir do seu index privado usando `[tool.uv.sources]`:
+
+```toml
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+### Exemplo completo
+
+```toml
+[project]
+name = "my-crew-project"
+version = "0.1.0"
+requires-python = ">=3.10,<=3.13"
+dependencies = [
+    "crewai[tools]>=0.100.1,<1.0.0",
+    "my-private-package>=1.2.0",
+]
+
+[tool.crewai]
+type = "crew"
+
+[[tool.uv.index]]
+name = "my-private-registry"
+url = "https://pkgs.dev.azure.com/my-org/_packaging/my-feed/pypi/simple/"
+explicit = true
+
+[tool.uv.sources]
+my-private-package = { index = "my-private-registry" }
+```
+
+Após atualizar o `pyproject.toml`, regenere seu arquivo lock:
+
+```bash
+uv lock
+```
+
+<Warning>
+  Sempre faça commit do `uv.lock` atualizado junto com as alterações no `pyproject.toml`.
+  O arquivo lock é obrigatório para implantação — veja [Preparar para Implantação](/pt-BR/enterprise/guides/prepare-for-deployment).
+</Warning>
+
+## Passo 2: Configurar Credenciais de Autenticação
+
+O UV autentica em indexes privados usando variáveis de ambiente que seguem uma convenção de nomenclatura
+baseada no nome do index que você definiu no `pyproject.toml`:
+
+```
+UV_INDEX_{UPPER_NAME}_USERNAME
+UV_INDEX_{UPPER_NAME}_PASSWORD
+```
+
+Onde `{UPPER_NAME}` é o nome do seu index convertido para **maiúsculas** com **hifens substituídos por underscores**.
+
+Por exemplo, um index chamado `my-private-registry` usa:
+
+| Variável | Valor |
+|----------|-------|
+| `UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME` | Seu nome de usuário ou nome do token do registro |
+| `UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD` | Sua senha ou token/PAT do registro |
+
+<Warning>
+  Essas variáveis de ambiente **devem** ser adicionadas pelas configurações de **Variáveis de Ambiente** do CrewAI AMP —
+  globalmente ou no nível da implantação. Elas não podem ser definidas em arquivos `.env` ou codificadas no seu projeto.
+
+  Veja [Configurar Variáveis de Ambiente no AMP](#configurar-variáveis-de-ambiente-no-amp) abaixo.
+</Warning>
+
+## Referência de Provedores de Registro
+
+A tabela abaixo mostra o formato da URL de index e os valores de credenciais para provedores de registro comuns.
+Substitua os valores de exemplo pelos detalhes reais da sua organização e feed.
+
+| Provedor | URL do Index | Usuário | Senha |
+|----------|-------------|---------|-------|
+| **Azure DevOps Artifacts** | `https://pkgs.dev.azure.com/{org}/_packaging/{feed}/pypi/simple/` | Qualquer string não vazia (ex: `token`) | Personal Access Token (PAT) com escopo Packaging Read |
+| **GitHub Packages** | `https://pypi.pkg.github.com/{owner}/simple/` | Nome de usuário do GitHub | Personal Access Token (classic) com escopo `read:packages` |
+| **GitLab Package Registry** | `https://gitlab.com/api/v4/projects/{project_id}/packages/pypi/simple/` | `__token__` | Project ou Personal Access Token com escopo `read_api` |
+| **AWS CodeArtifact** | Use a URL de `aws codeartifact get-repository-endpoint` | `aws` | Token de `aws codeartifact get-authorization-token` |
+| **Google Artifact Registry** | `https://{region}-python.pkg.dev/{project}/{repo}/simple/` | `_json_key_base64` | Chave de conta de serviço codificada em Base64 |
+| **JFrog Artifactory** | `https://{instance}.jfrog.io/artifactory/api/pypi/{repo}/simple/` | Nome de usuário ou email | Chave API ou token de identidade |
+| **Auto-hospedado (devpi, Nexus, etc.)** | URL da API simple do seu registro | Nome de usuário do registro | Senha do registro |
+
+<Tip>
+  Para **AWS CodeArtifact**, o token de autorização expira periodicamente.
+  Você precisará atualizar o valor de `UV_INDEX_*_PASSWORD` quando ele expirar.
+  Considere automatizar isso no seu pipeline de CI/CD.
+</Tip>
+
+## Configurar Variáveis de Ambiente no AMP
+
+As credenciais do registro privado devem ser configuradas como variáveis de ambiente no CrewAI AMP.
+Você tem duas opções:
+
+<Tabs>
+  <Tab title="Interface Web">
+    1. Faça login no [CrewAI AMP](https://app.crewai.com)
+    2. Navegue até sua automação
+    3. Abra a aba **Environment Variables**
+    4. Adicione cada variável (`UV_INDEX_*_USERNAME` e `UV_INDEX_*_PASSWORD`) com seu valor
+
+    Veja o passo [Deploy para AMP — Definir Variáveis de Ambiente](/pt-BR/enterprise/guides/deploy-to-amp#definir-as-variáveis-de-ambiente) para detalhes.
+  </Tab>
+  <Tab title="Implantação via CLI">
+    Adicione as variáveis ao seu arquivo `.env` local antes de executar `crewai deploy create`.
+    A CLI as transferirá com segurança para a plataforma:
+
+    ```bash
+    # .env
+    OPENAI_API_KEY=sk-...
+    UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+    UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat-here
+    ```
+
+    ```bash
+    crewai deploy create
+    ```
+  </Tab>
+</Tabs>
+
+<Warning>
+  **Nunca** faça commit de credenciais no seu repositório. Use variáveis de ambiente do AMP para todos os segredos.
+  O arquivo `.env` deve estar listado no `.gitignore`.
+</Warning>
+
+Para atualizar credenciais em uma implantação existente, veja [Atualizar Seu Crew — Variáveis de Ambiente](/pt-BR/enterprise/guides/update-crew).
+
+## Como Tudo se Conecta
+
+Quando o CrewAI AMP faz o build da sua automação, o fluxo de resolução funciona assim:
+
+<Steps>
+  <Step title="Build inicia">
+    O AMP busca seu repositório e lê o `pyproject.toml` e o `uv.lock`.
+  </Step>
+  <Step title="UV resolve dependências">
+    O UV lê `[tool.uv.sources]` para determinar de qual index cada pacote deve vir.
+  </Step>
+  <Step title="UV autentica">
+    Para cada index privado, o UV busca `UV_INDEX_{NAME}_USERNAME` e `UV_INDEX_{NAME}_PASSWORD`
+    nas variáveis de ambiente que você configurou no AMP.
+  </Step>
+  <Step title="Pacotes são instalados">
+    O UV baixa e instala todos os pacotes — tanto públicos (do PyPI) quanto privados (do seu registro).
+  </Step>
+  <Step title="Automação executa">
+    Seu crew ou flow inicia com todas as dependências disponíveis.
+  </Step>
+</Steps>
+
+## Solução de Problemas
+
+### Erros de Autenticação Durante o Build
+
+**Sintoma**: Build falha com `401 Unauthorized` ou `403 Forbidden` ao resolver um pacote privado.
+
+**Verifique**:
+- Os nomes das variáveis de ambiente `UV_INDEX_*` correspondem exatamente ao nome do seu index (maiúsculas, hifens -> underscores)
+- As credenciais estão definidas nas variáveis de ambiente do AMP, não apenas em um `.env` local
+- Seu token/PAT tem as permissões de leitura necessárias para o feed de pacotes
+- O token não expirou (especialmente relevante para AWS CodeArtifact)
+
+### Pacote Não Encontrado
+
+**Sintoma**: `No matching distribution found for my-private-package`.
+
+**Verifique**:
+- A URL do index no `pyproject.toml` termina com `/simple/`
+- A entrada `[tool.uv.sources]` mapeia o nome correto do pacote para o nome correto do index
+- O pacote está realmente publicado no seu registro privado
+- Execute `uv lock` localmente com as mesmas credenciais para verificar se a resolução funciona
+
+### Conflitos no Arquivo Lock
+
+**Sintoma**: `uv lock` falha ou produz resultados inesperados após adicionar um index privado.
+
+**Solução**: Defina as credenciais localmente e regenere:
+
+```bash
+export UV_INDEX_MY_PRIVATE_REGISTRY_USERNAME=token
+export UV_INDEX_MY_PRIVATE_REGISTRY_PASSWORD=your-pat
+uv lock
+```
+
+Em seguida, faça commit do `uv.lock` atualizado.
+
+## Guias Relacionados
+
+<CardGroup cols={3}>
+  <Card title="Preparar para Implantação" icon="clipboard-check" href="/pt-BR/enterprise/guides/prepare-for-deployment">
+    Verifique a estrutura do projeto e as dependências antes de implantar.
+  </Card>
+  <Card title="Deploy para AMP" icon="rocket" href="/pt-BR/enterprise/guides/deploy-to-amp">
+    Implante seu crew ou flow e configure variáveis de ambiente.
+  </Card>
+  <Card title="Atualizar Seu Crew" icon="arrows-rotate" href="/pt-BR/enterprise/guides/update-crew">
+    Atualize variáveis de ambiente e envie alterações para uma implantação em execução.
+  </Card>
+</CardGroup>

docs/pt-BR/guides/migration/migrating-from-langgraph.mdx

@@ -0,0 +1,518 @@
+---
+title: "Migrando do LangGraph para o CrewAI: um guia prático para engenheiros"
+description: Se você já construiu com LangGraph, saiba como portar rapidamente seus projetos para o CrewAI
+icon: switch
+mode: "wide"
+---
+
+Você construiu agentes com LangGraph. Já lutou com o `StateGraph`, ligou arestas condicionais e depurou dicionários de estado às 2 da manhã. Funciona — mas, em algum momento, você começou a se perguntar se existe um caminho melhor para produção.
+
+Existe. **CrewAI Flows** entrega o mesmo poder — orquestração orientada a eventos, roteamento condicional, estado compartilhado — com muito menos boilerplate e um modelo mental que se alinha a como você realmente pensa sobre fluxos de trabalho de IA em múltiplas etapas.
+
+Este artigo apresenta os conceitos principais lado a lado, mostra comparações reais de código e demonstra por que o CrewAI Flows é o framework que você vai querer usar a seguir.
+
+---
+
+## A Mudança de Modelo Mental
+
+LangGraph pede que você pense em **grafos**: nós, arestas e dicionários de estado. Todo workflow é um grafo direcionado em que você conecta explicitamente as transições entre as etapas de computação. É poderoso, mas a abstração traz overhead — especialmente quando o seu fluxo é fundamentalmente sequencial com alguns pontos de decisão.
+
+CrewAI Flows pede que você pense em **eventos**: métodos que iniciam, métodos que escutam resultados e métodos que roteiam a execução. A topologia do workflow emerge de anotações com decorators, em vez de construção explícita do grafo. Isso não é apenas açúcar sintático — muda como você projeta, lê e mantém seus pipelines.
+
+Veja o mapeamento principal:
+
+| Conceito no LangGraph | Equivalente no CrewAI Flows |
+| --- | --- |
+| `StateGraph` class | `Flow` class |
+| `add_node()` | Methods decorated with `@start`, `@listen` |
+| `add_edge()` / `add_conditional_edges()` | `@listen()` / `@router()` decorators |
+| `TypedDict` state | Pydantic `BaseModel` state |
+| `START` / `END` constants | `@start()` decorator / natural method return |
+| `graph.compile()` | `flow.kickoff()` |
+| Checkpointer / persistence | Built-in memory (LanceDB-backed) |
+
+Vamos ver como isso fica na prática.
+
+---
+
+## Demo 1: Um Pipeline Sequencial Simples
+
+Imagine que você está construindo um pipeline que recebe um tema, pesquisa, escreve um resumo e formata a saída. Veja como cada framework lida com isso.
+
+### Abordagem com LangGraph
+
+```python
+from typing import TypedDict
+from langgraph.graph import StateGraph, START, END
+
+class ResearchState(TypedDict):
+    topic: str
+    raw_research: str
+    summary: str
+    formatted_output: str
+
+def research_topic(state: ResearchState) -> dict:
+    # Call an LLM or search API
+    result = llm.invoke(f"Research the topic: {state['topic']}")
+    return {"raw_research": result}
+
+def write_summary(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Summarize this research:\n{state['raw_research']}"
+    )
+    return {"summary": result}
+
+def format_output(state: ResearchState) -> dict:
+    result = llm.invoke(
+        f"Format this summary as a polished article section:\n{state['summary']}"
+    )
+    return {"formatted_output": result}
+
+# Build the graph
+graph = StateGraph(ResearchState)
+graph.add_node("research", research_topic)
+graph.add_node("summarize", write_summary)
+graph.add_node("format", format_output)
+
+graph.add_edge(START, "research")
+graph.add_edge("research", "summarize")
+graph.add_edge("summarize", "format")
+graph.add_edge("format", END)
+
+# Compile and run
+app = graph.compile()
+result = app.invoke({"topic": "quantum computing advances in 2026"})
+print(result["formatted_output"])
+```
+
+Você define funções, registra-as como nós e conecta manualmente cada transição. Para uma sequência simples como essa, há muita cerimônia.
+
+### Abordagem com CrewAI Flows
+
+```python
+from crewai import LLM, Agent, Crew, Process, Task
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ResearchState(BaseModel):
+    topic: str = ""
+    raw_research: str = ""
+    summary: str = ""
+    formatted_output: str = ""
+
+class ResearchFlow(Flow[ResearchState]):
+    @start()
+    def research_topic(self):
+        # Option 1: Direct LLM call
+        result = llm.call(f"Research the topic: {self.state.topic}")
+        self.state.raw_research = result
+        return result
+
+    @listen(research_topic)
+    def write_summary(self, research_output):
+        # Option 2: A single agent
+        summarizer = Agent(
+            role="Research Summarizer",
+            goal="Produce concise, accurate summaries of research content",
+            backstory="You are an expert at distilling complex research into clear, "
+            "digestible summaries.",
+            llm=llm,
+            verbose=True,
+        )
+        result = summarizer.kickoff(
+            f"Summarize this research:\n{self.state.raw_research}"
+        )
+        self.state.summary = str(result)
+        return self.state.summary
+
+    @listen(write_summary)
+    def format_output(self, summary_output):
+        # Option 3: a complete crew (with one or more agents)
+        formatter = Agent(
+            role="Content Formatter",
+            goal="Transform research summaries into polished, publication-ready article sections",
+            backstory="You are a skilled editor with expertise in structuring and "
+            "presenting technical content for a general audience.",
+            llm=llm,
+            verbose=True,
+        )
+        format_task = Task(
+            description=f"Format this summary as a polished article section:\n{self.state.summary}",
+            expected_output="A well-structured, polished article section ready for publication.",
+            agent=formatter,
+        )
+        crew = Crew(
+            agents=[formatter],
+            tasks=[format_task],
+            process=Process.sequential,
+            verbose=True,
+        )
+        result = crew.kickoff()
+        self.state.formatted_output = str(result)
+        return self.state.formatted_output
+
+# Run the flow
+flow = ResearchFlow()
+flow.state.topic = "quantum computing advances in 2026"
+result = flow.kickoff()
+print(flow.state.formatted_output)
+
+```
+
+Repare a diferença: nada de construção de grafo, de ligação de arestas, nem de etapa de compilação. A ordem de execução é declarada exatamente onde a lógica vive. `@start()` marca o ponto de entrada, e `@listen(method_name)` encadeia as etapas. O estado é um modelo Pydantic de verdade, com segurança de tipos, validação e auto-complete na IDE.
+
+---
+
+## Demo 2: Roteamento Condicional
+
+Aqui é que fica interessante. Digamos que você está construindo um pipeline de conteúdo que roteia para diferentes caminhos de processamento com base no tipo de conteúdo detectado.
+
+### Abordagem com LangGraph
+
+```python
+from typing import TypedDict, Literal
+from langgraph.graph import StateGraph, START, END
+
+class ContentState(TypedDict):
+    input_text: str
+    content_type: str
+    result: str
+
+def classify_content(state: ContentState) -> dict:
+    content_type = llm.invoke(
+        f"Classify this content as 'technical', 'creative', or 'business':\n{state['input_text']}"
+    )
+    return {"content_type": content_type.strip().lower()}
+
+def process_technical(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as technical doc:\n{state['input_text']}")
+    return {"result": result}
+
+def process_creative(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as creative writing:\n{state['input_text']}")
+    return {"result": result}
+
+def process_business(state: ContentState) -> dict:
+    result = llm.invoke(f"Process as business content:\n{state['input_text']}")
+    return {"result": result}
+
+# Routing function
+def route_content(state: ContentState) -> Literal["technical", "creative", "business"]:
+    return state["content_type"]
+
+# Build the graph
+graph = StateGraph(ContentState)
+graph.add_node("classify", classify_content)
+graph.add_node("technical", process_technical)
+graph.add_node("creative", process_creative)
+graph.add_node("business", process_business)
+
+graph.add_edge(START, "classify")
+graph.add_conditional_edges(
+    "classify",
+    route_content,
+    {
+        "technical": "technical",
+        "creative": "creative",
+        "business": "business",
+    }
+)
+graph.add_edge("technical", END)
+graph.add_edge("creative", END)
+graph.add_edge("business", END)
+
+app = graph.compile()
+result = app.invoke({"input_text": "Explain how TCP handshakes work"})
+```
+
+Você precisa de uma função de roteamento separada, de um mapeamento explícito de arestas condicionais e de arestas de término para cada ramificação. A lógica de roteamento fica desacoplada do nó que produz a decisão.
+
+### Abordagem com CrewAI Flows
+
+```python
+from crewai import LLM, Agent
+from crewai.flow.flow import Flow, listen, router, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class ContentState(BaseModel):
+    input_text: str = ""
+    content_type: str = ""
+    result: str = ""
+
+class ContentFlow(Flow[ContentState]):
+    @start()
+    def classify_content(self):
+        self.state.content_type = (
+            llm.call(
+                f"Classify this content as 'technical', 'creative', or 'business':\n"
+                f"{self.state.input_text}"
+            )
+            .strip()
+            .lower()
+        )
+        return self.state.content_type
+
+    @router(classify_content)
+    def route_content(self, classification):
+        if classification == "technical":
+            return "process_technical"
+        elif classification == "creative":
+            return "process_creative"
+        else:
+            return "process_business"
+
+    @listen("process_technical")
+    def handle_technical(self):
+        agent = Agent(
+            role="Technical Writer",
+            goal="Produce clear, accurate technical documentation",
+            backstory="You are an expert technical writer who specializes in "
+            "explaining complex technical concepts precisely.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as technical doc:\n{self.state.input_text}")
+        )
+
+    @listen("process_creative")
+    def handle_creative(self):
+        agent = Agent(
+            role="Creative Writer",
+            goal="Craft engaging and imaginative creative content",
+            backstory="You are a talented creative writer with a flair for "
+            "compelling storytelling and vivid expression.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as creative writing:\n{self.state.input_text}")
+        )
+
+    @listen("process_business")
+    def handle_business(self):
+        agent = Agent(
+            role="Business Writer",
+            goal="Produce professional, results-oriented business content",
+            backstory="You are an experienced business writer who communicates "
+            "strategy and value clearly to professional audiences.",
+            llm=llm,
+            verbose=True,
+        )
+        self.state.result = str(
+            agent.kickoff(f"Process as business content:\n{self.state.input_text}")
+        )
+
+flow = ContentFlow()
+flow.state.input_text = "Explain how TCP handshakes work"
+flow.kickoff()
+print(flow.state.result)
+
+```
+
+O decorator `@router()` transforma um método em um ponto de decisão. Ele retorna uma string que corresponde a um listener — sem dicionários de mapeamento, sem funções de roteamento separadas. A lógica de ramificação parece um `if` em Python porque *é* um.
+
+---
+
+## Demo 3: Integrando Crews de Agentes de IA em Flows
+
+É aqui que o verdadeiro poder do CrewAI aparece. Flows não servem apenas para encadear chamadas de LLM — elas orquestram **Crews** completas de agentes autônomos. Isso é algo para o qual o LangGraph simplesmente não tem um equivalente nativo.
+
+```python
+from crewai import Agent, Task, Crew
+from crewai.flow.flow import Flow, listen, start
+from pydantic import BaseModel
+
+class ArticleState(BaseModel):
+    topic: str = ""
+    research: str = ""
+    draft: str = ""
+    final_article: str = ""
+
+class ArticleFlow(Flow[ArticleState]):
+
+    @start()
+    def run_research_crew(self):
+        """A full Crew of agents handles research."""
+        researcher = Agent(
+            role="Senior Research Analyst",
+            goal=f"Produce comprehensive research on: {self.state.topic}",
+            backstory="You're a veteran analyst known for thorough, "
+                       "well-sourced research reports.",
+            llm="gpt-4o"
+        )
+
+        research_task = Task(
+            description=f"Research '{self.state.topic}' thoroughly. "
+                        "Cover key trends, data points, and expert opinions.",
+            expected_output="A detailed research brief with sources.",
+            agent=researcher
+        )
+
+        crew = Crew(agents=[researcher], tasks=[research_task])
+        result = crew.kickoff()
+        self.state.research = result.raw
+        return result.raw
+
+    @listen(run_research_crew)
+    def run_writing_crew(self, research_output):
+        """A different Crew handles writing."""
+        writer = Agent(
+            role="Technical Writer",
+            goal="Write a compelling article based on provided research.",
+            backstory="You turn complex research into engaging, clear prose.",
+            llm="gpt-4o"
+        )
+
+        editor = Agent(
+            role="Senior Editor",
+            goal="Review and polish articles for publication quality.",
+            backstory="20 years of editorial experience at top tech publications.",
+            llm="gpt-4o"
+        )
+
+        write_task = Task(
+            description=f"Write an article based on this research:\n{self.state.research}",
+            expected_output="A well-structured draft article.",
+            agent=writer
+        )
+
+        edit_task = Task(
+            description="Review, fact-check, and polish the draft article.",
+            expected_output="A publication-ready article.",
+            agent=editor
+        )
+
+        crew = Crew(agents=[writer, editor], tasks=[write_task, edit_task])
+        result = crew.kickoff()
+        self.state.final_article = result.raw
+        return result.raw
+
+# Run the full pipeline
+flow = ArticleFlow()
+flow.state.topic = "The Future of Edge AI"
+flow.kickoff()
+print(flow.state.final_article)
+```
+
+Este é o insight-chave: **Flows fornecem a camada de orquestração, e Crews fornecem a camada de inteligência.** Cada etapa em um Flow pode subir uma equipe completa de agentes colaborativos, cada um com seus próprios papéis, objetivos e ferramentas. Você obtém fluxo de controle estruturado e previsível *e* colaboração autônoma de agentes — o melhor dos dois mundos.
+
+No LangGraph, alcançar algo similar significa implementar manualmente protocolos de comunicação entre agentes, loops de chamada de ferramentas e lógica de delegação dentro das funções dos nós. É possível, mas é encanamento que você constrói do zero todas as vezes.
+
+---
+
+## Demo 4: Execução Paralela e Sincronização
+
+Pipelines do mundo real frequentemente precisam dividir o trabalho e juntar os resultados. O CrewAI Flows lida com isso de forma elegante com os operadores `and_` e `or_`.
+
+```python
+from crewai import LLM
+from crewai.flow.flow import Flow, and_, listen, start
+from pydantic import BaseModel
+
+llm = LLM(model="openai/gpt-5.2")
+
+class AnalysisState(BaseModel):
+    topic: str = ""
+    market_data: str = ""
+    tech_analysis: str = ""
+    competitor_intel: str = ""
+    final_report: str = ""
+
+class ParallelAnalysisFlow(Flow[AnalysisState]):
+    @start()
+    def start_method(self):
+        pass
+
+    @listen(start_method)
+    def gather_market_data(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def run_tech_analysis(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(start_method)
+    def gather_competitor_intel(self):
+        # Your agentic or deterministic code
+        pass
+
+    @listen(and_(gather_market_data, run_tech_analysis, gather_competitor_intel))
+    def synthesize_report(self):
+        # Your agentic or deterministic code
+        pass
+
+flow = ParallelAnalysisFlow()
+flow.state.topic = "AI-powered developer tools"
+flow.kickoff()
+
+```
+
+Vários decorators `@start()` disparam em paralelo. O combinador `and_()` no decorator `@listen` garante que `synthesize_report` só execute depois que *todos os três* métodos upstream forem concluídos. Também existe `or_()` para quando você quer prosseguir assim que *qualquer* tarefa upstream terminar.
+
+No LangGraph, você precisaria construir um padrão fan-out/fan-in com ramificações paralelas, um nó de sincronização e uma mesclagem de estado cuidadosa — tudo conectado explicitamente por arestas.
+
+---
+
+## Por que CrewAI Flows em Produção
+
+Além de uma sintaxe mais limpa, Flows entrega várias vantagens críticas para produção:
+
+**Persistência de estado integrada.** O estado do Flow é respaldado pelo LanceDB, o que significa que seus workflows podem sobreviver a falhas, ser retomados e acumular conhecimento entre execuções. No LangGraph, você precisa configurar um checkpointer separado.
+
+**Gerenciamento de estado com segurança de tipos.** Modelos Pydantic oferecem validação, serialização e suporte de IDE prontos para uso. Estados `TypedDict` do LangGraph não validam em runtime.
+
+**Orquestração de agentes de primeira classe.** Crews são um primitivo nativo. Você define agentes com papéis, objetivos, histórias e ferramentas — e eles colaboram de forma autônoma dentro do envelope estruturado de um Flow. Não é preciso reinventar a coordenação multiagente.
+
+**Modelo mental mais simples.** Decorators declaram intenção. `@start` significa "comece aqui". `@listen(x)` significa "execute depois de x". `@router(x)` significa "decida para onde ir depois de x". O código lê como o workflow que ele descreve.
+
+**Integração com CLI.** Execute flows com `crewai run`. Sem etapa de compilação separada, sem serialização de grafo. Seu Flow é uma classe Python, e ele roda como tal.
+
+---
+
+## Cheat Sheet de Migração
+
+Se você está com uma base de código LangGraph e quer migrar para o CrewAI Flows, aqui vai um guia prático de conversão:
+
+1. **Mapeie seu estado.** Converta seu `TypedDict` para um `BaseModel` do Pydantic. Adicione valores padrão para todos os campos.
+2. **Converta nós em métodos.** Cada função de `add_node` vira um método na sua subclasse de `Flow`. Substitua leituras `state["field"]` por `self.state.field`.
+3. **Substitua arestas por decorators.** `add_edge(START, "first_node")` vira `@start()` no primeiro método. A sequência `add_edge("a", "b")` vira `@listen(a)` no método `b`.
+4. **Substitua arestas condicionais por `@router`.** A função de roteamento e o mapeamento do `add_conditional_edges()` viram um único método `@router()` que retorna a string de rota.
+5. **Troque compile + invoke por kickoff.** Remova `graph.compile()`. Chame `flow.kickoff()`.
+6. **Considere onde as Crews se encaixam.** Qualquer nó com lógica complexa de agentes em múltiplas etapas é um candidato a extração para uma Crew. É aqui que você verá a maior melhoria de qualidade.
+
+---
+
+## Primeiros Passos
+
+Instale o CrewAI e crie o scaffold de um novo projeto Flow:
+
+```bash
+pip install crewai
+crewai create flow my_first_flow
+cd my_first_flow
+```
+
+Isso gera uma estrutura de projeto com uma classe Flow pronta para edição, arquivos de configuração e um `pyproject.toml` com `type = "flow"` já definido. Execute com:
+
+```bash
+crewai run
+```
+
+A partir daí, adicione seus agentes, conecte seus listeners e publique.
+
+---
+
+## Considerações Finais
+
+O LangGraph ensinou ao ecossistema que workflows de IA precisam de estrutura. Essa foi uma lição importante. Mas o CrewAI Flows pega essa lição e a entrega de um jeito mais rápido de escrever, mais fácil de ler e mais poderoso em produção — especialmente quando seus workflows envolvem múltiplos agentes colaborando.
+
+Se você está construindo algo além de uma cadeia de agente único, dê uma olhada séria no Flows. O modelo baseado em decorators, a integração nativa com Crews e o gerenciamento de estado embutido significam menos tempo com encanamento e mais tempo nos problemas que importam.
+
+Comece com `crewai create flow`. Você não vai olhar para trás.

docs/pt-BR/learn/human-feedback-in-flows.mdx

@@ -98,33 +98,43 @@ def handle_feedback(self, result):
 Quando você especifica `emit`, o decorador se torna um roteador. O feedback livre do humano é interpretado por um LLM e mapeado para um dos outcomes especificados:
 
 ```python Code
-@start()
-@human_feedback(
-    message="Você aprova este conteúdo para publicação?",
-    emit=["approved", "rejected", "needs_revision"],
-    llm="gpt-4o-mini",
-    default_outcome="needs_revision",
-)
-def review_content(self):
-    return "Rascunho do post do blog aqui..."
+from crewai.flow.flow import Flow, start, listen, or_
+from crewai.flow.human_feedback import human_feedback
 
-@listen("approved")
-def publish(self, result):
-    print(f"Publicando! Usuário disse: {result.feedback}")
+class ReviewFlow(Flow):
+    @start()
+    def generate_content(self):
+        return "Rascunho do post do blog aqui..."
 
-@listen("rejected")
-def discard(self, result):
-    print(f"Descartando. Motivo: {result.feedback}")
+    @human_feedback(
+        message="Você aprova este conteúdo para publicação?",
+        emit=["approved", "rejected", "needs_revision"],
+        llm="gpt-4o-mini",
+        default_outcome="needs_revision",
+    )
+    @listen(or_("generate_content", "needs_revision"))
+    def review_content(self):
+        return "Rascunho do post do blog aqui..."
 
-@listen("needs_revision")
-def revise(self, result):
-    print(f"Revisando baseado em: {result.feedback}")
+    @listen("approved")
+    def publish(self, result):
+        print(f"Publicando! Usuário disse: {result.feedback}")
+
+    @listen("rejected")
+    def discard(self, result):
+        print(f"Descartando. Motivo: {result.feedback}")
 ```
 
+Quando o humano diz algo como "precisa de mais detalhes", o LLM mapeia para `"needs_revision"`, que dispara `review_content` novamente via `or_()` — criando um loop de revisão. O loop continua até que o outcome seja `"approved"` ou `"rejected"`.
+
 <Tip>
 O LLM usa saídas estruturadas (function calling) quando disponível para garantir que a resposta seja um dos seus outcomes especificados. Isso torna o roteamento confiável e previsível.
 </Tip>
 
+<Warning>
+Um método `@start()` só executa uma vez no início do flow. Se você precisa de um loop de revisão, separe o método start do método de revisão e use `@listen(or_("trigger", "revision_outcome"))` no método de revisão para habilitar o self-loop.
+</Warning>
+
 ## HumanFeedbackResult
 
 O dataclass `HumanFeedbackResult` contém todas as informações sobre uma interação de feedback humano:
@@ -193,116 +203,162 @@ Aqui está um exemplo completo implementando um fluxo de revisão e aprovação
 <CodeGroup>
 
 ```python Code
-from crewai.flow.flow import Flow, start, listen
+from crewai.flow.flow import Flow, start, listen, or_
 from crewai.flow.human_feedback import human_feedback, HumanFeedbackResult
 from pydantic import BaseModel
 
 
 class ContentState(BaseModel):
-    topic: str = ""
     draft: str = ""
-    final_content: str = ""
     revision_count: int = 0
+    status: str = "pending"
 
 
 class ContentApprovalFlow(Flow[ContentState]):
-    """Um flow que gera conteúdo e obtém aprovação humana."""
+    """Um flow que gera conteúdo e faz loop até o humano aprovar."""
 
     @start()
-    def get_topic(self):
-        self.state.topic = input("Sobre qual tópico devo escrever? ")
-        return self.state.topic
-
-    @listen(get_topic)
-    def generate_draft(self, topic):
-        # Em uso real, isso chamaria um LLM
-        self.state.draft = f"# {topic}\n\nEste é um rascunho sobre {topic}..."
+    def generate_draft(self):
+        self.state.draft = "# IA Segura\n\nEste é um rascunho sobre IA Segura..."
         return self.state.draft
 
-    @listen(generate_draft)
     @human_feedback(
-        message="Por favor, revise este rascunho. Responda 'approved', 'rejected', ou forneça feedback de revisão:",
+        message="Por favor, revise este rascunho. Aprove, rejeite ou descreva o que precisa mudar:",
         emit=["approved", "rejected", "needs_revision"],
         llm="gpt-4o-mini",
         default_outcome="needs_revision",
     )
-    def review_draft(self, draft):
-        return draft
+    @listen(or_("generate_draft", "needs_revision"))
+    def review_draft(self):
+        self.state.revision_count += 1
+        return f"{self.state.draft} (v{self.state.revision_count})"
 
     @listen("approved")
     def publish_content(self, result: HumanFeedbackResult):
-        self.state.final_content = result.output
-        print("\n✅ Conteúdo aprovado e publicado!")
-        print(f"Comentário do revisor: {result.feedback}")
+        self.state.status = "published"
+        print(f"Conteúdo aprovado e publicado! Revisor disse: {result.feedback}")
         return "published"
 
     @listen("rejected")
     def handle_rejection(self, result: HumanFeedbackResult):
-        print("\n❌ Conteúdo rejeitado")
-        print(f"Motivo: {result.feedback}")
+        self.state.status = "rejected"
+        print(f"Conteúdo rejeitado. Motivo: {result.feedback}")
         return "rejected"
 
-    @listen("needs_revision")
-    def revise_content(self, result: HumanFeedbackResult):
-        self.state.revision_count += 1
-        print(f"\n📝 Revisão #{self.state.revision_count} solicitada")
-        print(f"Feedback: {result.feedback}")
 
-        # Em um flow real, você pode voltar para generate_draft
-        # Para este exemplo, apenas reconhecemos
-        return "revision_requested"
-
-
-# Executar o flow
 flow = ContentApprovalFlow()
 result = flow.kickoff()
-print(f"\nFlow concluído. Revisões solicitadas: {flow.state.revision_count}")
+print(f"\nFlow finalizado. Status: {flow.state.status}, Revisões: {flow.state.revision_count}")
 ```
 
 ```text Output
-Sobre qual tópico devo escrever? Segurança em IA
+==================================================
+OUTPUT FOR REVIEW:
+==================================================
+# IA Segura
+
+Este é um rascunho sobre IA Segura... (v1)
+==================================================
+
+Por favor, revise este rascunho. Aprove, rejeite ou descreva o que precisa mudar:
+(Press Enter to skip, or type your feedback)
+
+Your feedback: Preciso de mais detalhes sobre segurança em IA.
 
 ==================================================
 OUTPUT FOR REVIEW:
 ==================================================
-# Segurança em IA
+# IA Segura
 
-Este é um rascunho sobre Segurança em IA...
+Este é um rascunho sobre IA Segura... (v2)
 ==================================================
 
-Por favor, revise este rascunho. Responda 'approved', 'rejected', ou forneça feedback de revisão:
+Por favor, revise este rascunho. Aprove, rejeite ou descreva o que precisa mudar:
 (Press Enter to skip, or type your feedback)
 
 Your feedback: Parece bom, aprovado!
 
-✅ Conteúdo aprovado e publicado!
-Comentário do revisor: Parece bom, aprovado!
+Conteúdo aprovado e publicado! Revisor disse: Parece bom, aprovado!
 
-Flow concluído. Revisões solicitadas: 0
+Flow finalizado. Status: published, Revisões: 2
 ```
 
 </CodeGroup>
 
 ## Combinando com Outros Decoradores
 
-O decorador `@human_feedback` funciona com outros decoradores de flow. Coloque-o como o decorador mais interno (mais próximo da função):
+O decorador `@human_feedback` funciona com `@start()`, `@listen()` e `or_()`. Ambas as ordens de decoradores funcionam — o framework propaga atributos em ambas as direções — mas os padrões recomendados são:
 
 ```python Code
-# Correto: @human_feedback é o mais interno (mais próximo da função)
+# Revisão única no início do flow (sem self-loop)
 @start()
-@human_feedback(message="Revise isto:")
+@human_feedback(message="Revise isto:", emit=["approved", "rejected"], llm="gpt-4o-mini")
 def my_start_method(self):
     return "content"
 
+# Revisão linear em um listener (sem self-loop)
 @listen(other_method)
-@human_feedback(message="Revise isto também:")
+@human_feedback(message="Revise isto também:", emit=["good", "bad"], llm="gpt-4o-mini")
 def my_listener(self, data):
     return f"processed: {data}"
+
+# Self-loop: revisão que pode voltar para revisões
+@human_feedback(message="Aprovar ou revisar?", emit=["approved", "revise"], llm="gpt-4o-mini")
+@listen(or_("upstream_method", "revise"))
+def review_with_loop(self):
+    return "content for review"
 ```
 
-<Tip>
-Coloque `@human_feedback` como o decorador mais interno (último/mais próximo da função) para que ele envolva o método diretamente e possa capturar o valor de retorno antes de passar para o sistema de flow.
-</Tip>
+### Padrão de self-loop
+
+Para criar um loop de revisão, o método de revisão deve escutar **ambos** um gatilho upstream e seu próprio outcome de revisão usando `or_()`:
+
+```python Code
+@start()
+def generate(self):
+    return "initial draft"
+
+@human_feedback(
+    message="Aprovar ou solicitar alterações?",
+    emit=["revise", "approved"],
+    llm="gpt-4o-mini",
+    default_outcome="approved",
+)
+@listen(or_("generate", "revise"))
+def review(self):
+    return "content"
+
+@listen("approved")
+def publish(self):
+    return "published"
+```
+
+Quando o outcome é `"revise"`, o flow roteia de volta para `review` (porque ele escuta `"revise"` via `or_()`). Quando o outcome é `"approved"`, o flow continua para `publish`. Isso funciona porque o engine de flow isenta roteadores da regra "fire once", permitindo que eles re-executem em cada iteração do loop.
+
+### Roteadores encadeados
+
+Um listener disparado pelo outcome de um roteador pode ser ele mesmo um roteador:
+
+```python Code
+@start()
+@human_feedback(message="Primeira revisão:", emit=["approved", "rejected"], llm="gpt-4o-mini")
+def draft(self):
+    return "draft content"
+
+@listen("approved")
+@human_feedback(message="Revisão final:", emit=["publish", "revise"], llm="gpt-4o-mini")
+def final_review(self, prev):
+    return "final content"
+
+@listen("publish")
+def on_publish(self, prev):
+    return "published"
+```
+
+### Limitações
+
+- **Métodos `@start()` executam uma vez**: Um método `@start()` não pode fazer self-loop. Se você precisa de um ciclo de revisão, use um método `@start()` separado como ponto de entrada e coloque o `@human_feedback` em um método `@listen()`.
+- **Sem `@start()` + `@listen()` no mesmo método**: Esta é uma restrição do framework de Flow. Um método é ou um ponto de início ou um listener, não ambos.
 
 ## Melhores Práticas
 
@@ -516,9 +572,9 @@ class ContentPipeline(Flow):
     @start()
     @human_feedback(
         message="Aprova este conteúdo para publicação?",
-        emit=["approved", "rejected", "needs_revision"],
+        emit=["approved", "rejected"],
         llm="gpt-4o-mini",
-        default_outcome="needs_revision",
+        default_outcome="rejected",
         provider=SlackNotificationProvider("#content-reviews"),
     )
     def generate_content(self):
@@ -534,11 +590,6 @@ class ContentPipeline(Flow):
         print(f"Arquivado. Motivo: {result.feedback}")
         return {"status": "archived"}
 
-    @listen("needs_revision")
-    def queue_revision(self, result):
-        print(f"Na fila para revisão: {result.feedback}")
-        return {"status": "revision_needed"}
-
 
 # Iniciando o flow (vai pausar e aguardar resposta do Slack)
 def start_content_pipeline():
@@ -594,22 +645,22 @@ Com o tempo, o humano vê saídas pré-revisadas progressivamente melhores porqu
 ```python Code
 class ArticleReviewFlow(Flow):
     @start()
+    def generate_article(self):
+        return self.crew.kickoff(inputs={"topic": "AI Safety"}).raw
+
     @human_feedback(
-        message="Review this article draft:",
+        message="Revise este rascunho do artigo:",
         emit=["approved", "needs_revision"],
         llm="gpt-4o-mini",
         learn=True,  # enable HITL learning
     )
-    def generate_article(self):
-        return self.crew.kickoff(inputs={"topic": "AI Safety"}).raw
+    @listen(or_("generate_article", "needs_revision"))
+    def review_article(self):
+        return self.last_human_feedback.output if self.last_human_feedback else "article draft"
 
     @listen("approved")
     def publish(self):
         print(f"Publishing: {self.last_human_feedback.output}")
-
-    @listen("needs_revision")
-    def revise(self):
-        print("Revising based on feedback...")
 ```
 
 **Primeira execução**: O humano vê a saída bruta e diz "Sempre inclua citações para afirmações factuais." A lição é destilada e armazenada na memória.

docs/pt-BR/learn/llm-connections.mdx

@@ -7,7 +7,7 @@ mode: "wide"
 
 ## Conecte o CrewAI a LLMs
 
-O CrewAI utiliza o LiteLLM para conectar-se a uma grande variedade de Modelos de Linguagem (LLMs). Essa integração proporciona grande versatilidade, permitindo que você utilize modelos de inúmeros provedores por meio de uma interface simples e unificada.
+O CrewAI conecta-se a LLMs por meio de integrações nativas via SDK para os provedores mais populares (OpenAI, Anthropic, Google Gemini, Azure e AWS Bedrock), e usa o LiteLLM como alternativa flexível para todos os demais provedores.
 
 <Note>
     Por padrão, o CrewAI usa o modelo `gpt-4o-mini`. Isso é determinado pela variável de ambiente `OPENAI_MODEL_NAME`, que tem como padrão "gpt-4o-mini" se não for definida.
@@ -40,6 +40,14 @@ O LiteLLM oferece suporte a uma ampla gama de provedores, incluindo, mas não se
 
 Para uma lista completa e sempre atualizada dos provedores suportados, consulte a [documentação de Provedores do LiteLLM](https://docs.litellm.ai/docs/providers).
 
+<Info>
+  Para usar qualquer provedor não coberto por uma integração nativa, adicione o LiteLLM como dependência ao seu projeto:
+  ```bash
+  uv add 'crewai[litellm]'
+  ```
+  Provedores nativos (OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock) usam seus próprios extras de SDK — consulte os [Exemplos de Configuração de Provedores](/pt-BR/concepts/llms#exemplos-de-configuração-de-provedores).
+</Info>
+
 ## Alterando a LLM
 
 Para utilizar uma LLM diferente com seus agentes CrewAI, você tem várias opções:

docs/pt-BR/tools/automation/composiotool.mdx

@@ -11,84 +11,53 @@ mode: "wide"
 Composio é uma plataforma de integração que permite conectar seus agentes de IA a mais de 250 ferramentas. Os principais recursos incluem:
 
 - **Autenticação de Nível Empresarial**: Suporte integrado para OAuth, Chaves de API, JWT com atualização automática de token
-- **Observabilidade Completa**: Logs detalhados de uso das ferramentas, registros de execução, e muito mais
+- **Observabilidade Completa**: Logs detalhados de uso das ferramentas, carimbos de data/hora de execução e muito mais
 
 ## Instalação
 
 Para incorporar as ferramentas Composio em seu projeto, siga as instruções abaixo:
 
 ```shell
-pip install composio-crewai
+pip install composio composio-crewai
 pip install crewai
 ```
 
-Após a conclusão da instalação, execute `composio login` ou exporte sua chave de API do composio como `COMPOSIO_API_KEY`. Obtenha sua chave de API Composio [aqui](https://app.composio.dev)
+Após concluir a instalação, defina sua chave de API do Composio como `COMPOSIO_API_KEY`. Obtenha sua chave de API do Composio [aqui](https://platform.composio.dev)
 
 ## Exemplo
 
-O exemplo a seguir demonstra como inicializar a ferramenta e executar uma ação do github:
+O exemplo a seguir demonstra como inicializar a ferramenta e executar uma ação do GitHub:
 
-1. Inicialize o conjunto de ferramentas Composio
+1. Inicialize o Composio com o Provider do CrewAI
 
 ```python Code
-from composio_crewai import ComposioToolSet, App, Action
+from composio_crewai import ComposioProvider
+from composio import Composio
 from crewai import Agent, Task, Crew
 
-toolset = ComposioToolSet()
+composio = Composio(provider=ComposioProvider())
 ```
 
-2. Conecte sua conta do GitHub
+2. Crie uma nova sessão Composio e recupere as ferramentas
 <CodeGroup>
-```shell CLI
-composio add github
-```
-```python Code
-request = toolset.initiate_connection(app=App.GITHUB)
-print(f"Open this URL to authenticate: {request.redirectUrl}")
+```python
+session = composio.create(
+    user_id="your-user-id",
+    toolkits=["gmail", "github"] # optional, default is all toolkits
+)
+tools = session.tools()
 ```
+Leia mais sobre sessões e gerenciamento de usuários [aqui](https://docs.composio.dev/docs/configuring-sessions)
 </CodeGroup>
 
-3. Obtenha ferramentas
+3. Autenticação manual dos usuários
 
-- Recuperando todas as ferramentas de um app (não recomendado em produção):
+O Composio autentica automaticamente os usuários durante a sessão de chat do agente. No entanto, você também pode autenticar o usuário manualmente chamando o método `authorize`.
 ```python Code
-tools = toolset.get_tools(apps=[App.GITHUB])
+connection_request = session.authorize("github")
+print(f"Open this URL to authenticate: {connection_request.redirect_url}")
 ```
 
-- Filtrando ferramentas com base em tags:
-```python Code
-tag = "users"
-
-filtered_action_enums = toolset.find_actions_by_tags(
-    App.GITHUB,
-    tags=[tag], 
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-
-- Filtrando ferramentas com base no caso de uso:
-```python Code
-use_case = "Star a repository on GitHub"
-
-filtered_action_enums = toolset.find_actions_by_use_case(
-    App.GITHUB, use_case=use_case, advanced=False
-)
-
-tools = toolset.get_tools(actions=filtered_action_enums)
-```
-<Tip>Defina `advanced` como True para obter ações para casos de uso complexos</Tip>
-
-- Usando ferramentas específicas:
-
-Neste exemplo, usaremos a ação `GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER` do app GitHub.
-```python Code
-tools = toolset.get_tools(
-    actions=[Action.GITHUB_STAR_A_REPOSITORY_FOR_THE_AUTHENTICATED_USER]
-)
-```
-Saiba mais sobre como filtrar ações [aqui](https://docs.composio.dev/patterns/tools/use-tools/use-specific-actions)
-
 4. Defina o agente
 
 ```python Code
@@ -116,4 +85,4 @@ crew = Crew(agents=[crewai_agent], tasks=[task])
 crew.kickoff()
 ```
 
-* Uma lista mais detalhada de ferramentas pode ser encontrada [aqui](https://app.composio.dev)
+* Uma lista mais detalhada de ferramentas pode ser encontrada [aqui](https://docs.composio.dev/toolkits)

lib/crewai-files/pyproject.toml

@@ -8,8 +8,8 @@ authors = [
 ]
 requires-python = ">=3.10, <3.14"
 dependencies = [
-    "Pillow~=10.4.0",
-    "pypdf~=4.0.0",
+    "Pillow~=12.1.1",
+    "pypdf~=6.7.5",
     "python-magic>=0.4.27",
     "aiocache~=0.12.3",
     "aiofiles~=24.1.0",

lib/crewai-files/src/crewai_files/__init__.py

@@ -152,4 +152,4 @@ __all__ = [
     "wrap_file_source",
 ]
 
-__version__ = "1.9.3"
+__version__ = "1.10.1"

lib/crewai-tools/pyproject.toml

@@ -8,12 +8,10 @@ authors = [
 ]
 requires-python = ">=3.10, <3.14"
 dependencies = [
-    "lancedb~=0.5.4",
     "pytube~=15.0.0",
     "requests~=2.32.5",
     "docker~=7.1.0",
-    "crewai==1.9.3",
-    "lancedb~=0.5.4",
+    "crewai==1.10.1",
     "tiktoken~=0.8.0",
     "beautifulsoup4~=4.13.4",
     "python-docx~=1.2.0",

lib/crewai-tools/src/crewai_tools/__init__.py

@@ -291,4 +291,4 @@ __all__ = [
     "ZapierActionTools",
 ]
 
-__version__ = "1.9.3"
+__version__ = "1.10.1"

lib/crewai-tools/src/crewai_tools/tools/brave_search_tool/brave_search_tool.py

@@ -10,6 +10,7 @@ from pydantic import BaseModel, Field
 from pydantic.types import StringConstraints
 import requests
 
+
 load_dotenv()
 
 

lib/crewai-tools/src/crewai_tools/tools/multion_tool/example.py

@@ -1,7 +1,7 @@
 import os
 
 from crewai import Agent, Crew, Task
-from multion_tool import MultiOnTool # type: ignore[import-not-found]
+from multion_tool import MultiOnTool  # type: ignore[import-not-found]
 
 
 os.environ["OPENAI_API_KEY"] = "Your Key"

lib/crewai-tools/src/crewai_tools/tools/stagehand_tool/example.py

@@ -17,11 +17,11 @@ Usage:
 
 import os
 
+from crewai import Agent, Crew, Process, Task
 from crewai.utilities.printer import Printer
 from dotenv import load_dotenv
 from stagehand.schemas import AvailableModel  # type: ignore[import-untyped]
 
-from crewai import Agent, Crew, Process, Task
 from crewai_tools import StagehandTool
 
 

lib/crewai-tools/tool.specs.json

@@ -20117,18 +20117,6 @@
       "humanized_name": "Web Automation Tool",
       "init_params_schema": {
         "$defs": {
-          "AvailableModel": {
-            "enum": [
-              "gpt-4o",
-              "gpt-4o-mini",
-              "claude-3-5-sonnet-latest",
-              "claude-3-7-sonnet-latest",
-              "computer-use-preview",
-              "gemini-2.0-flash"
-            ],
-            "title": "AvailableModel",
-            "type": "string"
-          },
           "EnvVar": {
             "properties": {
               "default": {
@@ -20206,17 +20194,6 @@
             "default": null,
             "title": "Model Api Key"
           },
-          "model_name": {
-            "anyOf": [
-              {
-                "$ref": "#/$defs/AvailableModel"
-              },
-              {
-                "type": "null"
-              }
-            ],
-            "default": "claude-3-7-sonnet-latest"
-          },
           "project_id": {
             "anyOf": [
               {

lib/crewai/pyproject.toml

@@ -21,7 +21,7 @@ dependencies = [
     "opentelemetry-exporter-otlp-proto-http~=1.34.0",
     # Data Handling
     "chromadb~=1.1.0",
-    "tokenizers~=0.20.3",
+    "tokenizers>=0.21,<1",
     "openpyxl~=3.1.5",
     # Authentication and Security
     "python-dotenv~=1.1.1",
@@ -38,10 +38,11 @@ dependencies = [
     "json5~=0.10.0",
     "portalocker~=2.7.0",
     "pydantic-settings~=2.10.1",
+    "httpx~=0.28.1",
     "mcp~=1.26.0",
     "uv~=0.9.13",
     "aiosqlite~=0.21.0",
-    "lancedb>=0.4.0",
+    "lancedb>=0.29.2",
 ]
 
 [project.urls]
@@ -52,7 +53,7 @@ Repository = "https://github.com/crewAIInc/crewAI"
 
 [project.optional-dependencies]
 tools = [
-    "crewai-tools==1.9.3",
+    "crewai-tools==1.10.1",
 ]
 embeddings = [
     "tiktoken~=0.8.0"
@@ -65,7 +66,7 @@ openpyxl = [
 ]
 mem0 = ["mem0ai~=0.1.94"]
 docling = [
-    "docling~=2.63.0",
+    "docling~=2.75.0",
 ]
 qdrant = [
     "qdrant-client[fastembed]~=1.14.3",
@@ -87,7 +88,7 @@ bedrock = [
     "boto3~=1.40.45",
 ]
 google-genai = [
-    "google-genai~=1.49.0",
+    "google-genai~=1.65.0",
 ]
 azure-ai-inference = [
     "azure-ai-inference~=1.0.0b9",

lib/crewai/src/crewai/__init__.py

@@ -4,13 +4,13 @@ import urllib.request
 import warnings
 
 from crewai.agent.core import Agent
+from crewai.agent.planning_config import PlanningConfig
 from crewai.crew import Crew
 from crewai.crews.crew_output import CrewOutput
 from crewai.flow.flow import Flow
 from crewai.knowledge.knowledge import Knowledge
 from crewai.llm import LLM
 from crewai.llms.base_llm import BaseLLM
-from crewai.memory.unified_memory import Memory
 from crewai.process import Process
 from crewai.task import Task
 from crewai.tasks.llm_guardrail import LLMGuardrail
@@ -41,7 +41,7 @@ def _suppress_pydantic_deprecation_warnings() -> None:
 
 _suppress_pydantic_deprecation_warnings()
 
-__version__ = "1.9.3"
+__version__ = "1.10.1"
 _telemetry_submitted = False
 
 
@@ -72,6 +72,25 @@ def _track_install_async() -> None:
 
 
 _track_install_async()
+
+_LAZY_IMPORTS: dict[str, tuple[str, str]] = {
+    "Memory": ("crewai.memory.unified_memory", "Memory"),
+}
+
+
+def __getattr__(name: str) -> Any:
+    """Lazily import heavy modules (e.g. Memory → lancedb) on first access."""
+    if name in _LAZY_IMPORTS:
+        module_path, attr = _LAZY_IMPORTS[name]
+        import importlib
+
+        mod = importlib.import_module(module_path)
+        val = getattr(mod, attr)
+        globals()[name] = val
+        return val
+    raise AttributeError(f"module 'crewai' has no attribute {name!r}")
+
+
 __all__ = [
     "LLM",
     "Agent",
@@ -82,6 +101,7 @@ __all__ = [
     "Knowledge",
     "LLMGuardrail",
     "Memory",
+    "PlanningConfig",
     "Process",
     "Task",
     "TaskOutput",

lib/crewai/src/crewai/a2a/utils/agent_card.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import MutableMapping
+import concurrent.futures
 from functools import lru_cache
 import ssl
 import time
@@ -138,14 +139,17 @@ def fetch_agent_card(
         ttl_hash = int(time.time() // cache_ttl)
         return _fetch_agent_card_cached(endpoint, auth_hash, timeout, ttl_hash)
 
-    loop = asyncio.new_event_loop()
-    asyncio.set_event_loop(loop)
+    coro = afetch_agent_card(endpoint=endpoint, auth=auth, timeout=timeout)
     try:
-        return loop.run_until_complete(
-            afetch_agent_card(endpoint=endpoint, auth=auth, timeout=timeout)
-        )
-    finally:
-        loop.close()
+        asyncio.get_running_loop()
+        has_running_loop = True
+    except RuntimeError:
+        has_running_loop = False
+
+    if has_running_loop:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            return pool.submit(asyncio.run, coro).result()
+    return asyncio.run(coro)
 
 
 async def afetch_agent_card(
@@ -203,14 +207,17 @@ def _fetch_agent_card_cached(
     """Cached sync version of fetch_agent_card."""
     auth = _auth_store.get(auth_hash)
 
-    loop = asyncio.new_event_loop()
-    asyncio.set_event_loop(loop)
+    coro = _afetch_agent_card_impl(endpoint=endpoint, auth=auth, timeout=timeout)
     try:
-        return loop.run_until_complete(
-            _afetch_agent_card_impl(endpoint=endpoint, auth=auth, timeout=timeout)
-        )
-    finally:
-        loop.close()
+        asyncio.get_running_loop()
+        has_running_loop = True
+    except RuntimeError:
+        has_running_loop = False
+
+    if has_running_loop:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            return pool.submit(asyncio.run, coro).result()
+    return asyncio.run(coro)
 
 
 @cached(ttl=300, serializer=PickleSerializer())  # type: ignore[untyped-decorator]

lib/crewai/src/crewai/a2a/utils/delegation.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import base64
 from collections.abc import AsyncIterator, Callable, MutableMapping
+import concurrent.futures
 from contextlib import asynccontextmanager
 import logging
 from typing import TYPE_CHECKING, Any, Final, Literal
@@ -194,56 +195,43 @@ def execute_a2a_delegation(
 
     Returns:
         TaskStateResult with status, result/error, history, and agent_card.
-
-    Raises:
-        RuntimeError: If called from an async context with a running event loop.
     """
+    coro = aexecute_a2a_delegation(
+        endpoint=endpoint,
+        auth=auth,
+        timeout=timeout,
+        task_description=task_description,
+        context=context,
+        context_id=context_id,
+        task_id=task_id,
+        reference_task_ids=reference_task_ids,
+        metadata=metadata,
+        extensions=extensions,
+        conversation_history=conversation_history,
+        agent_id=agent_id,
+        agent_role=agent_role,
+        agent_branch=agent_branch,
+        response_model=response_model,
+        turn_number=turn_number,
+        updates=updates,
+        from_task=from_task,
+        from_agent=from_agent,
+        skill_id=skill_id,
+        client_extensions=client_extensions,
+        transport=transport,
+        accepted_output_modes=accepted_output_modes,
+        input_files=input_files,
+    )
     try:
         asyncio.get_running_loop()
-        raise RuntimeError(
-            "execute_a2a_delegation() cannot be called from an async context. "
-            "Use 'await aexecute_a2a_delegation()' instead."
-        )
-    except RuntimeError as e:
-        if "no running event loop" not in str(e).lower():
-            raise
+        has_running_loop = True
+    except RuntimeError:
+        has_running_loop = False
 
-    loop = asyncio.new_event_loop()
-    asyncio.set_event_loop(loop)
-    try:
-        return loop.run_until_complete(
-            aexecute_a2a_delegation(
-                endpoint=endpoint,
-                auth=auth,
-                timeout=timeout,
-                task_description=task_description,
-                context=context,
-                context_id=context_id,
-                task_id=task_id,
-                reference_task_ids=reference_task_ids,
-                metadata=metadata,
-                extensions=extensions,
-                conversation_history=conversation_history,
-                agent_id=agent_id,
-                agent_role=agent_role,
-                agent_branch=agent_branch,
-                response_model=response_model,
-                turn_number=turn_number,
-                updates=updates,
-                from_task=from_task,
-                from_agent=from_agent,
-                skill_id=skill_id,
-                client_extensions=client_extensions,
-                transport=transport,
-                accepted_output_modes=accepted_output_modes,
-                input_files=input_files,
-            )
-        )
-    finally:
-        try:
-            loop.run_until_complete(loop.shutdown_asyncgens())
-        finally:
-            loop.close()
+    if has_running_loop:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            return pool.submit(asyncio.run, coro).result()
+    return asyncio.run(coro)
 
 
 async def aexecute_a2a_delegation(

lib/crewai/src/crewai/agent/core.py

@@ -8,11 +8,9 @@ import time
 from typing import (
     TYPE_CHECKING,
     Any,
-    Final,
     Literal,
     cast,
 )
-from urllib.parse import urlparse
 
 from pydantic import (
     BaseModel,
@@ -24,6 +22,7 @@ from pydantic import (
 )
 from typing_extensions import Self
 
+from crewai.agent.planning_config import PlanningConfig
 from crewai.agent.utils import (
     ahandle_knowledge_retrieval,
     apply_training_data,
@@ -61,16 +60,8 @@ from crewai.knowledge.knowledge import Knowledge
 from crewai.knowledge.source.base_knowledge_source import BaseKnowledgeSource
 from crewai.lite_agent_output import LiteAgentOutput
 from crewai.llms.base_llm import BaseLLM
-from crewai.mcp import (
-    MCPClient,
-    MCPServerConfig,
-    MCPServerHTTP,
-    MCPServerSSE,
-    MCPServerStdio,
-)
-from crewai.mcp.transports.http import HTTPTransport
-from crewai.mcp.transports.sse import SSETransport
-from crewai.mcp.transports.stdio import StdioTransport
+from crewai.mcp import MCPServerConfig
+from crewai.mcp.tool_resolver import MCPToolResolver
 from crewai.rag.embeddings.types import EmbedderConfig
 from crewai.security.fingerprint import Fingerprint
 from crewai.tools.agent_tools.agent_tools import AgentTools
@@ -111,18 +102,8 @@ if TYPE_CHECKING:
     from crewai.utilities.types import LLMMessage
 
 
-# MCP Connection timeout constants (in seconds)
-MCP_CONNECTION_TIMEOUT: Final[int] = 10
-MCP_TOOL_EXECUTION_TIMEOUT: Final[int] = 30
-MCP_DISCOVERY_TIMEOUT: Final[int] = 15
-MCP_MAX_RETRIES: Final[int] = 3
-
 _passthrough_exceptions: tuple[type[Exception], ...] = ()
 
-# Simple in-memory cache for MCP tool schemas (duration: 5 minutes)
-_mcp_schema_cache: dict[str, Any] = {}
-_cache_ttl: Final[int] = 300  # 5 minutes
-
 
 class Agent(BaseAgent):
     """Represents an agent in a system.
@@ -154,7 +135,7 @@ class Agent(BaseAgent):
     model_config = ConfigDict()
 
     _times_executed: int = PrivateAttr(default=0)
-    _mcp_clients: list[Any] = PrivateAttr(default_factory=list)
+    _mcp_resolver: MCPToolResolver | None = PrivateAttr(default=None)
     _last_messages: list[LLMMessage] = PrivateAttr(default_factory=list)
     max_execution_time: int | None = Field(
         default=None,
@@ -211,13 +192,23 @@ class Agent(BaseAgent):
         default="safe",
         description="Mode for code execution: 'safe' (using Docker) or 'unsafe' (direct execution).",
     )
-    reasoning: bool = Field(
+    planning_config: PlanningConfig | None = Field(
+        default=None,
+        description="Configuration for agent planning before task execution.",
+    )
+    planning: bool = Field(
         default=False,
         description="Whether the agent should reflect and create a plan before executing a task.",
     )
+    reasoning: bool = Field(
+        default=False,
+        description="[DEPRECATED: Use planning_config instead] Whether the agent should reflect and create a plan before executing a task.",
+        deprecated=True,
+    )
     max_reasoning_attempts: int | None = Field(
         default=None,
-        description="Maximum number of reasoning attempts before executing the task. If None, will try until ready.",
+        description="[DEPRECATED: Use planning_config.max_attempts instead] Maximum number of reasoning attempts before executing the task. If None, will try until ready.",
+        deprecated=True,
     )
     embedder: EmbedderConfig | None = Field(
         default=None,
@@ -284,8 +275,26 @@ class Agent(BaseAgent):
         if self.allow_code_execution:
             self._validate_docker_installation()
 
+        # Handle backward compatibility: convert reasoning=True to planning_config
+        if self.reasoning and self.planning_config is None:
+            import warnings
+
+            warnings.warn(
+                "The 'reasoning' parameter is deprecated. Use 'planning_config=PlanningConfig()' instead.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+            self.planning_config = PlanningConfig(
+                max_attempts=self.max_reasoning_attempts,
+            )
+
         return self
 
+    @property
+    def planning_enabled(self) -> bool:
+        """Check if planning is enabled for this agent."""
+        return self.planning_config is not None or self.planning
+
     def _setup_agent_executor(self) -> None:
         if not self.cache_handler:
             self.cache_handler = CacheHandler()
@@ -354,7 +363,11 @@ class Agent(BaseAgent):
             ValueError: If the max execution time is not a positive integer.
             RuntimeError: If the agent execution fails for other reasons.
         """
-        handle_reasoning(self, task)
+        # Only call handle_reasoning for legacy CrewAgentExecutor
+        # For AgentExecutor, planning is handled in AgentExecutor.generate_plan()
+        if self.executor_class is not AgentExecutor:
+            handle_reasoning(self, task)
+
         self._inject_date_to_task(task)
 
         if self.tools_handler:
@@ -384,10 +397,10 @@ class Agent(BaseAgent):
                 )
                 if unified_memory is not None:
                     query = task.description
-                    matches = unified_memory.recall(query, limit=10)
+                    matches = unified_memory.recall(query, limit=5)
                     if matches:
                         memory = "Relevant memories:\n" + "\n".join(
-                            f"- {m.record.content}" for m in matches
+                            m.format() for m in matches
                         )
                 if memory.strip() != "":
                     task_prompt += self.i18n.slice("memory").format(memory=memory)
@@ -592,7 +605,10 @@ class Agent(BaseAgent):
             ValueError: If the max execution time is not a positive integer.
             RuntimeError: If the agent execution fails for other reasons.
         """
-        handle_reasoning(self, task)
+        if self.executor_class is not AgentExecutor:
+            handle_reasoning(
+                self, task
+            )  # we need this till CrewAgentExecutor migrates to AgentExecutor
         self._inject_date_to_task(task)
 
         if self.tools_handler:
@@ -622,10 +638,10 @@ class Agent(BaseAgent):
                 )
                 if unified_memory is not None:
                     query = task.description
-                    matches = unified_memory.recall(query, limit=10)
+                    matches = unified_memory.recall(query, limit=5)
                     if matches:
                         memory = "Relevant memories:\n" + "\n".join(
-                            f"- {m.record.content}" for m in matches
+                            m.format() for m in matches
                         )
                 if memory.strip() != "":
                     task_prompt += self.i18n.slice("memory").format(memory=memory)
@@ -864,7 +880,11 @@ class Agent(BaseAgent):
                 respect_context_window=self.respect_context_window,
                 request_within_rpm_limit=rpm_limit_fn,
                 callbacks=[TokenCalcHandler(self._token_process)],
-                response_model=task.response_model if task else None,
+                response_model=(
+                    task.response_model or task.output_pydantic or task.output_json
+                )
+                if task
+                else None,
             )
 
     def _update_executor_parameters(
@@ -893,7 +913,11 @@ class Agent(BaseAgent):
         self.agent_executor.stop = stop_words
         self.agent_executor.tools_names = get_tool_names(tools)
         self.agent_executor.tools_description = render_text_description_and_args(tools)
-        self.agent_executor.response_model = task.response_model if task else None
+        self.agent_executor.response_model = (
+            (task.response_model or task.output_pydantic or task.output_json)
+            if task
+            else None
+        )
 
         self.agent_executor.tools_handler = self.tools_handler
         self.agent_executor.request_within_rpm_limit = rpm_limit_fn
@@ -926,544 +950,17 @@ class Agent(BaseAgent):
     def get_mcp_tools(self, mcps: list[str | MCPServerConfig]) -> list[BaseTool]:
         """Convert MCP server references/configs to CrewAI tools.
 
-        Supports both string references (backwards compatible) and structured
-        configuration objects (MCPServerStdio, MCPServerHTTP, MCPServerSSE).
-
-        Args:
-            mcps: List of MCP server references (strings) or configurations.
-
-        Returns:
-            List of BaseTool instances from MCP servers.
+        Delegates to :class:`~crewai.mcp.tool_resolver.MCPToolResolver`.
         """
-        all_tools = []
-        clients = []
-
-        for mcp_config in mcps:
-            if isinstance(mcp_config, str):
-                tools = self._get_mcp_tools_from_string(mcp_config)
-            else:
-                tools, client = self._get_native_mcp_tools(mcp_config)
-                if client:
-                    clients.append(client)
-
-            all_tools.extend(tools)
-
-        # Store clients for cleanup
-        self._mcp_clients.extend(clients)
-        return all_tools
+        self._cleanup_mcp_clients()
+        self._mcp_resolver = MCPToolResolver(agent=self, logger=self._logger)
+        return self._mcp_resolver.resolve(mcps)
 
     def _cleanup_mcp_clients(self) -> None:
         """Cleanup MCP client connections after task execution."""
-        if not self._mcp_clients:
-            return
-
-        async def _disconnect_all() -> None:
-            for client in self._mcp_clients:
-                if client and hasattr(client, "connected") and client.connected:
-                    await client.disconnect()
-
-        try:
-            asyncio.run(_disconnect_all())
-        except Exception as e:
-            self._logger.log("error", f"Error during MCP client cleanup: {e}")
-        finally:
-            self._mcp_clients.clear()
-
-    def _get_mcp_tools_from_string(self, mcp_ref: str) -> list[BaseTool]:
-        """Get tools from legacy string-based MCP references.
-
-        This method maintains backwards compatibility with string-based
-        MCP references (https://... and crewai-amp:...).
-
-        Args:
-            mcp_ref: String reference to MCP server.
-
-        Returns:
-            List of BaseTool instances.
-        """
-        if mcp_ref.startswith("crewai-amp:"):
-            return self._get_amp_mcp_tools(mcp_ref)
-        if mcp_ref.startswith("https://"):
-            return self._get_external_mcp_tools(mcp_ref)
-        return []
-
-    def _get_external_mcp_tools(self, mcp_ref: str) -> list[BaseTool]:
-        """Get tools from external HTTPS MCP server with graceful error handling."""
-        from crewai.tools.mcp_tool_wrapper import MCPToolWrapper
-
-        # Parse server URL and optional tool name
-        if "#" in mcp_ref:
-            server_url, specific_tool = mcp_ref.split("#", 1)
-        else:
-            server_url, specific_tool = mcp_ref, None
-
-        server_params = {"url": server_url}
-        server_name = self._extract_server_name(server_url)
-
-        try:
-            # Get tool schemas with timeout and error handling
-            tool_schemas = self._get_mcp_tool_schemas(server_params)
-
-            if not tool_schemas:
-                self._logger.log(
-                    "warning", f"No tools discovered from MCP server: {server_url}"
-                )
-                return []
-
-            tools = []
-            for tool_name, schema in tool_schemas.items():
-                # Skip if specific tool requested and this isn't it
-                if specific_tool and tool_name != specific_tool:
-                    continue
-
-                try:
-                    wrapper = MCPToolWrapper(
-                        mcp_server_params=server_params,
-                        tool_name=tool_name,
-                        tool_schema=schema,
-                        server_name=server_name,
-                    )
-                    tools.append(wrapper)
-                except Exception as e:
-                    self._logger.log(
-                        "warning",
-                        f"Failed to create MCP tool wrapper for {tool_name}: {e}",
-                    )
-                    continue
-
-            if specific_tool and not tools:
-                self._logger.log(
-                    "warning",
-                    f"Specific tool '{specific_tool}' not found on MCP server: {server_url}",
-                )
-
-            return cast(list[BaseTool], tools)
-
-        except Exception as e:
-            self._logger.log(
-                "warning", f"Failed to connect to MCP server {server_url}: {e}"
-            )
-            return []
-
-    def _get_native_mcp_tools(
-        self, mcp_config: MCPServerConfig
-    ) -> tuple[list[BaseTool], Any | None]:
-        """Get tools from MCP server using structured configuration.
-
-        This method creates an MCP client based on the configuration type,
-        connects to the server, discovers tools, applies filtering, and
-        returns wrapped tools along with the client instance for cleanup.
-
-        Args:
-            mcp_config: MCP server configuration (MCPServerStdio, MCPServerHTTP, or MCPServerSSE).
-
-        Returns:
-            Tuple of (list of BaseTool instances, MCPClient instance for cleanup).
-        """
-        from crewai.tools.base_tool import BaseTool
-        from crewai.tools.mcp_native_tool import MCPNativeTool
-
-        transport: StdioTransport | HTTPTransport | SSETransport
-        if isinstance(mcp_config, MCPServerStdio):
-            transport = StdioTransport(
-                command=mcp_config.command,
-                args=mcp_config.args,
-                env=mcp_config.env,
-            )
-            server_name = f"{mcp_config.command}_{'_'.join(mcp_config.args)}"
-        elif isinstance(mcp_config, MCPServerHTTP):
-            transport = HTTPTransport(
-                url=mcp_config.url,
-                headers=mcp_config.headers,
-                streamable=mcp_config.streamable,
-            )
-            server_name = self._extract_server_name(mcp_config.url)
-        elif isinstance(mcp_config, MCPServerSSE):
-            transport = SSETransport(
-                url=mcp_config.url,
-                headers=mcp_config.headers,
-            )
-            server_name = self._extract_server_name(mcp_config.url)
-        else:
-            raise ValueError(f"Unsupported MCP server config type: {type(mcp_config)}")
-
-        client = MCPClient(
-            transport=transport,
-            cache_tools_list=mcp_config.cache_tools_list,
-        )
-
-        async def _setup_client_and_list_tools() -> list[dict[str, Any]]:
-            """Async helper to connect and list tools in same event loop."""
-
-            try:
-                if not client.connected:
-                    await client.connect()
-
-                tools_list = await client.list_tools()
-
-                try:
-                    await client.disconnect()
-                    # Small delay to allow background tasks to finish cleanup
-                    # This helps prevent "cancel scope in different task" errors
-                    # when asyncio.run() closes the event loop
-                    await asyncio.sleep(0.1)
-                except Exception as e:
-                    self._logger.log("error", f"Error during disconnect: {e}")
-
-                return tools_list
-            except Exception as e:
-                if client.connected:
-                    await client.disconnect()
-                    await asyncio.sleep(0.1)
-                raise RuntimeError(
-                    f"Error during setup client and list tools: {e}"
-                ) from e
-
-        try:
-            try:
-                asyncio.get_running_loop()
-                import concurrent.futures
-
-                with concurrent.futures.ThreadPoolExecutor() as executor:
-                    future = executor.submit(
-                        asyncio.run, _setup_client_and_list_tools()
-                    )
-                    tools_list = future.result()
-            except RuntimeError:
-                try:
-                    tools_list = asyncio.run(_setup_client_and_list_tools())
-                except RuntimeError as e:
-                    error_msg = str(e).lower()
-                    if "cancel scope" in error_msg or "task" in error_msg:
-                        raise ConnectionError(
-                            "MCP connection failed due to event loop cleanup issues. "
-                            "This may be due to authentication errors or server unavailability."
-                        ) from e
-                except asyncio.CancelledError as e:
-                    raise ConnectionError(
-                        "MCP connection was cancelled. This may indicate an authentication "
-                        "error or server unavailability."
-                    ) from e
-
-            if mcp_config.tool_filter:
-                filtered_tools = []
-                for tool in tools_list:
-                    if callable(mcp_config.tool_filter):
-                        try:
-                            from crewai.mcp.filters import ToolFilterContext
-
-                            context = ToolFilterContext(
-                                agent=self,
-                                server_name=server_name,
-                                run_context=None,
-                            )
-                            if mcp_config.tool_filter(context, tool):  # type: ignore[call-arg, arg-type]
-                                filtered_tools.append(tool)
-                        except (TypeError, AttributeError):
-                            if mcp_config.tool_filter(tool):  # type: ignore[call-arg, arg-type]
-                                filtered_tools.append(tool)
-                    else:
-                        # Not callable - include tool
-                        filtered_tools.append(tool)
-                tools_list = filtered_tools
-
-            tools = []
-            for tool_def in tools_list:
-                tool_name = tool_def.get("name", "")
-                if not tool_name:
-                    continue
-
-                # Convert inputSchema to Pydantic model if present
-                args_schema = None
-                if tool_def.get("inputSchema"):
-                    args_schema = self._json_schema_to_pydantic(
-                        tool_name, tool_def["inputSchema"]
-                    )
-
-                tool_schema = {
-                    "description": tool_def.get("description", ""),
-                    "args_schema": args_schema,
-                }
-
-                try:
-                    native_tool = MCPNativeTool(
-                        mcp_client=client,
-                        tool_name=tool_name,
-                        tool_schema=tool_schema,
-                        server_name=server_name,
-                    )
-                    tools.append(native_tool)
-                except Exception as e:
-                    self._logger.log("error", f"Failed to create native MCP tool: {e}")
-                    continue
-
-            return cast(list[BaseTool], tools), client
-        except Exception as e:
-            if client.connected:
-                asyncio.run(client.disconnect())
-
-            raise RuntimeError(f"Failed to get native MCP tools: {e}") from e
-
-    def _get_amp_mcp_tools(self, amp_ref: str) -> list[BaseTool]:
-        """Get tools from CrewAI AMP MCP marketplace."""
-        # Parse: "crewai-amp:mcp-name" or "crewai-amp:mcp-name#tool_name"
-        amp_part = amp_ref.replace("crewai-amp:", "")
-        if "#" in amp_part:
-            mcp_name, specific_tool = amp_part.split("#", 1)
-        else:
-            mcp_name, specific_tool = amp_part, None
-
-        # Call AMP API to get MCP server URLs
-        mcp_servers = self._fetch_amp_mcp_servers(mcp_name)
-
-        tools = []
-        for server_config in mcp_servers:
-            server_ref = server_config["url"]
-            if specific_tool:
-                server_ref += f"#{specific_tool}"
-            server_tools = self._get_external_mcp_tools(server_ref)
-            tools.extend(server_tools)
-
-        return tools
-
-    @staticmethod
-    def _extract_server_name(server_url: str) -> str:
-        """Extract clean server name from URL for tool prefixing."""
-
-        parsed = urlparse(server_url)
-        domain = parsed.netloc.replace(".", "_")
-        path = parsed.path.replace("/", "_").strip("_")
-        return f"{domain}_{path}" if path else domain
-
-    def _get_mcp_tool_schemas(
-        self, server_params: dict[str, Any]
-    ) -> dict[str, dict[str, Any]]:
-        """Get tool schemas from MCP server for wrapper creation with caching."""
-        server_url = server_params["url"]
-
-        # Check cache first
-        cache_key = server_url
-        current_time = time.time()
-
-        if cache_key in _mcp_schema_cache:
-            cached_data, cache_time = _mcp_schema_cache[cache_key]
-            if current_time - cache_time < _cache_ttl:
-                self._logger.log(
-                    "debug", f"Using cached MCP tool schemas for {server_url}"
-                )
-                return cached_data  # type: ignore[no-any-return]
-
-        try:
-            schemas = asyncio.run(self._get_mcp_tool_schemas_async(server_params))
-
-            # Cache successful results
-            _mcp_schema_cache[cache_key] = (schemas, current_time)
-
-            return schemas
-        except Exception as e:
-            # Log warning but don't raise - this allows graceful degradation
-            self._logger.log(
-                "warning", f"Failed to get MCP tool schemas from {server_url}: {e}"
-            )
-            return {}
-
-    async def _get_mcp_tool_schemas_async(
-        self, server_params: dict[str, Any]
-    ) -> dict[str, dict[str, Any]]:
-        """Async implementation of MCP tool schema retrieval with timeouts and retries."""
-        server_url = server_params["url"]
-        return await self._retry_mcp_discovery(
-            self._discover_mcp_tools_with_timeout, server_url
-        )
-
-    async def _retry_mcp_discovery(
-        self, operation_func: Any, server_url: str
-    ) -> dict[str, dict[str, Any]]:
-        """Retry MCP discovery operation with exponential backoff, avoiding try-except in loop."""
-        last_error = None
-
-        for attempt in range(MCP_MAX_RETRIES):
-            # Execute single attempt outside try-except loop structure
-            result, error, should_retry = await self._attempt_mcp_discovery(
-                operation_func, server_url
-            )
-
-            # Success case - return immediately
-            if result is not None:
-                return result
-
-            # Non-retryable error - raise immediately
-            if not should_retry:
-                raise RuntimeError(error)
-
-            # Retryable error - continue with backoff
-            last_error = error
-            if attempt < MCP_MAX_RETRIES - 1:
-                wait_time = 2**attempt  # Exponential backoff
-                await asyncio.sleep(wait_time)
-
-        raise RuntimeError(
-            f"Failed to discover MCP tools after {MCP_MAX_RETRIES} attempts: {last_error}"
-        )
-
-    @staticmethod
-    async def _attempt_mcp_discovery(
-        operation_func: Any, server_url: str
-    ) -> tuple[dict[str, dict[str, Any]] | None, str, bool]:
-        """Attempt single MCP discovery operation and return (result, error_message, should_retry)."""
-        try:
-            result = await operation_func(server_url)
-            return result, "", False
-
-        except ImportError:
-            return (
-                None,
-                "MCP library not available. Please install with: pip install mcp",
-                False,
-            )
-
-        except asyncio.TimeoutError:
-            return (
-                None,
-                f"MCP discovery timed out after {MCP_DISCOVERY_TIMEOUT} seconds",
-                True,
-            )
-
-        except Exception as e:
-            error_str = str(e).lower()
-
-            # Classify errors as retryable or non-retryable
-            if "authentication" in error_str or "unauthorized" in error_str:
-                return None, f"Authentication failed for MCP server: {e!s}", False
-            if "connection" in error_str or "network" in error_str:
-                return None, f"Network connection failed: {e!s}", True
-            if "json" in error_str or "parsing" in error_str:
-                return None, f"Server response parsing error: {e!s}", True
-            return None, f"MCP discovery error: {e!s}", False
-
-    async def _discover_mcp_tools_with_timeout(
-        self, server_url: str
-    ) -> dict[str, dict[str, Any]]:
-        """Discover MCP tools with timeout wrapper."""
-        return await asyncio.wait_for(
-            self._discover_mcp_tools(server_url), timeout=MCP_DISCOVERY_TIMEOUT
-        )
-
-    async def _discover_mcp_tools(self, server_url: str) -> dict[str, dict[str, Any]]:
-        """Discover tools from MCP server with proper timeout handling."""
-        from mcp import ClientSession
-        from mcp.client.streamable_http import streamablehttp_client
-
-        async with streamablehttp_client(server_url) as (read, write, _):
-            async with ClientSession(read, write) as session:
-                # Initialize the connection with timeout
-                await asyncio.wait_for(
-                    session.initialize(), timeout=MCP_CONNECTION_TIMEOUT
-                )
-
-                # List available tools with timeout
-                tools_result = await asyncio.wait_for(
-                    session.list_tools(),
-                    timeout=MCP_DISCOVERY_TIMEOUT - MCP_CONNECTION_TIMEOUT,
-                )
-
-                schemas = {}
-                for tool in tools_result.tools:
-                    args_schema = None
-                    if hasattr(tool, "inputSchema") and tool.inputSchema:
-                        args_schema = self._json_schema_to_pydantic(
-                            sanitize_tool_name(tool.name), tool.inputSchema
-                        )
-
-                    schemas[sanitize_tool_name(tool.name)] = {
-                        "description": getattr(tool, "description", ""),
-                        "args_schema": args_schema,
-                    }
-                return schemas
-
-    def _json_schema_to_pydantic(
-        self, tool_name: str, json_schema: dict[str, Any]
-    ) -> type:
-        """Convert JSON Schema to Pydantic model for tool arguments.
-
-        Args:
-            tool_name: Name of the tool (used for model naming)
-            json_schema: JSON Schema dict with 'properties', 'required', etc.
-
-        Returns:
-            Pydantic BaseModel class
-        """
-        from pydantic import Field, create_model
-
-        properties = json_schema.get("properties", {})
-        required_fields = json_schema.get("required", [])
-
-        field_definitions: dict[str, Any] = {}
-
-        for field_name, field_schema in properties.items():
-            field_type = self._json_type_to_python(field_schema)
-            field_description = field_schema.get("description", "")
-
-            is_required = field_name in required_fields
-
-            if is_required:
-                field_definitions[field_name] = (
-                    field_type,
-                    Field(..., description=field_description),
-                )
-            else:
-                field_definitions[field_name] = (
-                    field_type | None,
-                    Field(default=None, description=field_description),
-                )
-
-        model_name = f"{tool_name.replace('-', '_').replace(' ', '_')}Schema"
-        return create_model(model_name, **field_definitions)  # type: ignore[no-any-return]
-
-    def _json_type_to_python(self, field_schema: dict[str, Any]) -> type:
-        """Convert JSON Schema type to Python type.
-
-        Args:
-            field_schema: JSON Schema field definition
-
-        Returns:
-            Python type
-        """
-
-        json_type = field_schema.get("type")
-
-        if "anyOf" in field_schema:
-            types: list[type] = []
-            for option in field_schema["anyOf"]:
-                if "const" in option:
-                    types.append(str)
-                else:
-                    types.append(self._json_type_to_python(option))
-            unique_types = list(set(types))
-            if len(unique_types) > 1:
-                result: Any = unique_types[0]
-                for t in unique_types[1:]:
-                    result = result | t
-                return result  # type: ignore[no-any-return]
-            return unique_types[0]
-
-        type_mapping: dict[str | None, type] = {
-            "string": str,
-            "number": float,
-            "integer": int,
-            "boolean": bool,
-            "array": list,
-            "object": dict,
-        }
-
-        return type_mapping.get(json_type, Any)
-
-    @staticmethod
-    def _fetch_amp_mcp_servers(mcp_name: str) -> list[dict[str, Any]]:
-        """Fetch MCP server configurations from CrewAI AMP API."""
-        # TODO: Implement AMP API call to "integrations/mcps" endpoint
-        # Should return list of server configs with URLs
-        return []
+        if self._mcp_resolver is not None:
+            self._mcp_resolver.cleanup()
+            self._mcp_resolver = None
 
     @staticmethod
     def get_multimodal_tools() -> Sequence[BaseTool]:
@@ -1695,11 +1192,15 @@ class Agent(BaseAgent):
         # Process platform apps and MCP tools
         if self.apps:
             platform_tools = self.get_platform_tools(self.apps)
-            if platform_tools and self.tools is not None:
+            if platform_tools:
+                if self.tools is None:
+                    self.tools = []
                 self.tools.extend(platform_tools)
         if self.mcps:
             mcps = self.get_mcp_tools(self.mcps)
-            if mcps and self.tools is not None:
+            if mcps:
+                if self.tools is None:
+                    self.tools = []
                 self.tools.extend(mcps)
 
         # Prepare tools
@@ -1712,7 +1213,8 @@ class Agent(BaseAgent):
 
             existing_names = {sanitize_tool_name(t.name) for t in raw_tools}
             raw_tools.extend(
-                mt for mt in create_memory_tools(agent_memory)
+                mt
+                for mt in create_memory_tools(agent_memory)
                 if sanitize_tool_name(mt.name) not in existing_names
             )
 
@@ -1802,11 +1304,11 @@ class Agent(BaseAgent):
                     ),
                 )
                 start_time = time.time()
-                matches = agent_memory.recall(formatted_messages, limit=10)
+                matches = agent_memory.recall(formatted_messages, limit=20)
                 memory_block = ""
                 if matches:
                     memory_block = "Relevant memories:\n" + "\n".join(
-                        f"- {m.record.content}" for m in matches
+                        m.format() for m in matches
                     )
                 if memory_block:
                     formatted_messages += "\n\n" + self.i18n.slice("memory").format(
@@ -1937,14 +1439,15 @@ class Agent(BaseAgent):
             if isinstance(messages, str):
                 input_str = messages
             else:
-                input_str = "\n".join(
-                    str(msg.get("content", "")) for msg in messages if msg.get("content")
-                ) or "User request"
-            raw = (
-                f"Input: {input_str}\n"
-                f"Agent: {self.role}\n"
-                f"Result: {output_text}"
-            )
+                input_str = (
+                    "\n".join(
+                        str(msg.get("content", ""))
+                        for msg in messages
+                        if msg.get("content")
+                    )
+                    or "User request"
+                )
+            raw = f"Input: {input_str}\nAgent: {self.role}\nResult: {output_text}"
             extracted = agent_memory.extract_memories(raw)
             if extracted:
                 agent_memory.remember_many(extracted)

lib/crewai/src/crewai/agent/planning_config.py

@@ -0,0 +1,83 @@
+from __future__ import annotations
+
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+
+class PlanningConfig(BaseModel):
+    """Configuration for agent planning/reasoning before task execution.
+
+    This allows users to customize the planning behavior including prompts,
+    iteration limits, and the LLM used for planning.
+
+    Note: To disable planning, don't pass a planning_config or set planning=False
+    on the Agent. The presence of a PlanningConfig enables planning.
+
+    Attributes:
+        max_attempts: Maximum number of planning refinement attempts.
+            If None, will continue until the agent indicates readiness.
+        max_steps: Maximum number of steps in the generated plan.
+        system_prompt: Custom system prompt for planning. Uses default if None.
+        plan_prompt: Custom prompt for creating the initial plan.
+        refine_prompt: Custom prompt for refining the plan.
+        llm: LLM to use for planning. Uses agent's LLM if None.
+
+    Example:
+        ```python
+        from crewai import Agent
+        from crewai.agent.planning_config import PlanningConfig
+
+        # Simple usage
+        agent = Agent(
+            role="Researcher",
+            goal="Research topics",
+            backstory="Expert researcher",
+            planning_config=PlanningConfig(),
+        )
+
+        # Customized planning
+        agent = Agent(
+            role="Researcher",
+            goal="Research topics",
+            backstory="Expert researcher",
+            planning_config=PlanningConfig(
+                max_attempts=3,
+                max_steps=10,
+                plan_prompt="Create a focused plan for: {description}",
+                llm="gpt-4o-mini",  # Use cheaper model for planning
+            ),
+        )
+        ```
+    """
+
+    max_attempts: int | None = Field(
+        default=None,
+        description=(
+            "Maximum number of planning refinement attempts. "
+            "If None, will continue until the agent indicates readiness."
+        ),
+    )
+    max_steps: int = Field(
+        default=20,
+        description="Maximum number of steps in the generated plan.",
+        ge=1,
+    )
+    system_prompt: str | None = Field(
+        default=None,
+        description="Custom system prompt for planning. Uses default if None.",
+    )
+    plan_prompt: str | None = Field(
+        default=None,
+        description="Custom prompt for creating the initial plan.",
+    )
+    refine_prompt: str | None = Field(
+        default=None,
+        description="Custom prompt for refining the plan.",
+    )
+    llm: str | Any | None = Field(
+        default=None,
+        description="LLM to use for planning. Uses agent's LLM if None.",
+    )
+
+    model_config = {"arbitrary_types_allowed": True}

lib/crewai/src/crewai/agent/utils.py

@@ -28,13 +28,20 @@ if TYPE_CHECKING:
 
 
 def handle_reasoning(agent: Agent, task: Task) -> None:
-    """Handle the reasoning process for an agent before task execution.
+    """Handle the reasoning/planning process for an agent before task execution.
+
+    This function checks if planning is enabled for the agent and, if so,
+    creates a plan that gets appended to the task description.
+
+    Note: This function is used by CrewAgentExecutor (legacy path).
+    For AgentExecutor, planning is handled in AgentExecutor.generate_plan().
 
     Args:
         agent: The agent performing the task.
         task: The task to execute.
     """
-    if not agent.reasoning:
+    # Check if planning is enabled using the planning_enabled property
+    if not getattr(agent, "planning_enabled", False):
         return
 
     try:
@@ -43,13 +50,13 @@ def handle_reasoning(agent: Agent, task: Task) -> None:
             AgentReasoningOutput,
         )
 
-        reasoning_handler = AgentReasoning(task=task, agent=agent)
-        reasoning_output: AgentReasoningOutput = (
-            reasoning_handler.handle_agent_reasoning()
+        planning_handler = AgentReasoning(agent=agent, task=task)
+        planning_output: AgentReasoningOutput = (
+            planning_handler.handle_agent_reasoning()
         )
-        task.description += f"\n\nReasoning Plan:\n{reasoning_output.plan.plan}"
+        task.description += f"\n\nPlanning:\n{planning_output.plan.plan}"
     except Exception as e:
-        agent._logger.log("error", f"Error during reasoning process: {e!s}")
+        agent._logger.log("error", f"Error during planning: {e!s}")
 
 
 def build_task_prompt_with_schema(task: Task, task_prompt: str, i18n: I18N) -> str:

lib/crewai/src/crewai/agents/agent_builder/base_agent.py

@@ -4,7 +4,8 @@ from abc import ABC, abstractmethod
 from collections.abc import Callable
 from copy import copy as shallow_copy
 from hashlib import md5
-from typing import Any, Literal
+import re
+from typing import Any, Final, Literal
 import uuid
 
 from pydantic import (
@@ -36,6 +37,11 @@ from crewai.utilities.rpm_controller import RPMController
 from crewai.utilities.string_utils import interpolate_only
 
 
+_SLUG_RE: Final[re.Pattern[str]] = re.compile(
+    r"^(?:crewai-amp:)?[a-zA-Z0-9][a-zA-Z0-9_-]*(?:#\w+)?$"
+)
+
+
 PlatformApp = Literal[
     "asana",
     "box",
@@ -197,7 +203,7 @@ class BaseAgent(BaseModel, ABC, metaclass=AgentMeta):
     )
     mcps: list[str | MCPServerConfig] | None = Field(
         default=None,
-        description="List of MCP server references. Supports 'https://server.com/path' for external servers and 'crewai-amp:mcp-name' for AMP marketplace. Use '#tool_name' suffix for specific tools.",
+        description="List of MCP server references. Supports 'https://server.com/path' for external servers and bare slugs like 'notion' for connected MCP integrations. Use '#tool_name' suffix for specific tools.",
     )
     memory: Any = Field(
         default=None,
@@ -276,14 +282,16 @@ class BaseAgent(BaseModel, ABC, metaclass=AgentMeta):
         validated_mcps: list[str | MCPServerConfig] = []
         for mcp in mcps:
             if isinstance(mcp, str):
-                if mcp.startswith(("https://", "crewai-amp:")):
+                if mcp.startswith("https://"):
+                    validated_mcps.append(mcp)
+                elif _SLUG_RE.match(mcp):
                     validated_mcps.append(mcp)
                 else:
                     raise ValueError(
-                        f"Invalid MCP reference: {mcp}. "
-                        "String references must start with 'https://' or 'crewai-amp:'"
+                        f"Invalid MCP reference: {mcp!r}. "
+                        "String references must be an 'https://' URL or a valid "
+                        "slug (e.g. 'notion', 'notion#search', 'crewai-amp:notion')."
                     )
-
             elif isinstance(mcp, (MCPServerConfig)):
                 validated_mcps.append(mcp)
             else:

lib/crewai/src/crewai/agents/agent_builder/base_agent_executor_mixin.py

@@ -30,7 +30,7 @@ class CrewAgentExecutorMixin:
         memory = getattr(self.agent, "memory", None) or (
             getattr(self.crew, "_memory", None) if self.crew else None
         )
-        if memory is None or not self.task:
+        if memory is None or not self.task or getattr(memory, "_read_only", False):
             return
         if (
             f"Action: {sanitize_tool_name('Delegate work to coworker')}"

lib/crewai/src/crewai/agents/cache/__init__.py

@@ -1,5 +1,4 @@
 from crewai.agents.cache.cache_handler import CacheHandler
 
 
-
 __all__ = ["CacheHandler"]

lib/crewai/src/crewai/agents/crew_agent_executor.py

@@ -6,7 +6,10 @@ and memory management.
 
 from __future__ import annotations
 
+import asyncio
 from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import inspect
 import logging
 from typing import TYPE_CHECKING, Any, Literal, cast
 
@@ -47,6 +50,7 @@ from crewai.utilities.agent_utils import (
     handle_unknown_error,
     has_reached_max_iterations,
     is_context_length_exceeded,
+    parse_tool_call_args,
     process_llm_response,
     track_delegation_if_needed,
 )
@@ -483,8 +487,8 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
             # No tools available, fall back to simple LLM call
             return self._invoke_loop_native_no_tools()
 
-        openai_tools, available_functions = convert_tools_to_openai_schema(
-            self.original_tools
+        openai_tools, available_functions, self._tool_name_mapping = (
+            convert_tools_to_openai_schema(self.original_tools)
         )
 
         while True:
@@ -685,30 +689,140 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
         Returns:
             AgentFinish if tool has result_as_answer=True, None otherwise.
         """
-        from datetime import datetime
-        import json
-
-        from crewai.events import crewai_event_bus
-        from crewai.events.types.tool_usage_events import (
-            ToolUsageErrorEvent,
-            ToolUsageFinishedEvent,
-            ToolUsageStartedEvent,
-        )
-
         if not tool_calls:
             return None
 
-        # Only process the FIRST tool call for sequential execution with reflection
-        tool_call = tool_calls[0]
+        parsed_calls = [
+            parsed
+            for tool_call in tool_calls
+            if (parsed := self._parse_native_tool_call(tool_call)) is not None
+        ]
+        if not parsed_calls:
+            return None
 
-        # Extract tool call info - handle OpenAI-style, Anthropic-style, and Gemini-style
+        original_tools_by_name: dict[str, Any] = dict(self._tool_name_mapping)
+
+        if len(parsed_calls) > 1:
+            has_result_as_answer_in_batch = any(
+                bool(
+                    original_tools_by_name.get(func_name)
+                    and getattr(
+                        original_tools_by_name.get(func_name), "result_as_answer", False
+                    )
+                )
+                for _, func_name, _ in parsed_calls
+            )
+            has_max_usage_count_in_batch = any(
+                bool(
+                    original_tools_by_name.get(func_name)
+                    and getattr(
+                        original_tools_by_name.get(func_name),
+                        "max_usage_count",
+                        None,
+                    )
+                    is not None
+                )
+                for _, func_name, _ in parsed_calls
+            )
+
+            # Preserve historical sequential behavior for result_as_answer batches.
+            # Also avoid threading around usage counters for max_usage_count tools.
+            if has_result_as_answer_in_batch or has_max_usage_count_in_batch:
+                logger.debug(
+                    "Skipping parallel native execution because batch includes result_as_answer or max_usage_count tool"
+                )
+            else:
+                execution_plan: list[
+                    tuple[str, str, str | dict[str, Any], Any | None]
+                ] = []
+                for call_id, func_name, func_args in parsed_calls:
+                    original_tool = original_tools_by_name.get(func_name)
+                    execution_plan.append(
+                        (call_id, func_name, func_args, original_tool)
+                    )
+
+                self._append_assistant_tool_calls_message(
+                    [
+                        (call_id, func_name, func_args)
+                        for call_id, func_name, func_args, _ in execution_plan
+                    ]
+                )
+
+                max_workers = min(8, len(execution_plan))
+                ordered_results: list[dict[str, Any] | None] = [None] * len(
+                    execution_plan
+                )
+                with ThreadPoolExecutor(max_workers=max_workers) as pool:
+                    futures = {
+                        pool.submit(
+                            self._execute_single_native_tool_call,
+                            call_id=call_id,
+                            func_name=func_name,
+                            func_args=func_args,
+                            available_functions=available_functions,
+                            original_tool=original_tool,
+                            should_execute=True,
+                        ): idx
+                        for idx, (
+                            call_id,
+                            func_name,
+                            func_args,
+                            original_tool,
+                        ) in enumerate(execution_plan)
+                    }
+                    for future in as_completed(futures):
+                        idx = futures[future]
+                        ordered_results[idx] = future.result()
+
+                for execution_result in ordered_results:
+                    if not execution_result:
+                        continue
+                    tool_finish = self._append_tool_result_and_check_finality(
+                        execution_result
+                    )
+                    if tool_finish:
+                        return tool_finish
+
+                reasoning_prompt = self._i18n.slice("post_tool_reasoning")
+                reasoning_message: LLMMessage = {
+                    "role": "user",
+                    "content": reasoning_prompt,
+                }
+                self.messages.append(reasoning_message)
+                return None
+
+        # Sequential behavior: process only first tool call, then force reflection.
+        call_id, func_name, func_args = parsed_calls[0]
+        self._append_assistant_tool_calls_message([(call_id, func_name, func_args)])
+
+        execution_result = self._execute_single_native_tool_call(
+            call_id=call_id,
+            func_name=func_name,
+            func_args=func_args,
+            available_functions=available_functions,
+            original_tool=original_tools_by_name.get(func_name),
+            should_execute=True,
+        )
+        tool_finish = self._append_tool_result_and_check_finality(execution_result)
+        if tool_finish:
+            return tool_finish
+
+        reasoning_prompt = self._i18n.slice("post_tool_reasoning")
+        reasoning_message = {
+            "role": "user",
+            "content": reasoning_prompt,
+        }
+        self.messages.append(reasoning_message)
+        return None
+
+    def _parse_native_tool_call(
+        self, tool_call: Any
+    ) -> tuple[str, str, str | dict[str, Any]] | None:
         if hasattr(tool_call, "function"):
-            # OpenAI-style: has .function.name and .function.arguments
             call_id = getattr(tool_call, "id", f"call_{id(tool_call)}")
             func_name = sanitize_tool_name(tool_call.function.name)
-            func_args = tool_call.function.arguments
-        elif hasattr(tool_call, "function_call") and tool_call.function_call:
-            # Gemini-style: has .function_call.name and .function_call.args
+            return call_id, func_name, tool_call.function.arguments
+        if hasattr(tool_call, "function_call") and tool_call.function_call:
             call_id = f"call_{id(tool_call)}"
             func_name = sanitize_tool_name(tool_call.function_call.name)
             func_args = (
@@ -716,13 +830,12 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 if tool_call.function_call.args
                 else {}
             )
-        elif hasattr(tool_call, "name") and hasattr(tool_call, "input"):
-            # Anthropic format: has .name and .input (ToolUseBlock)
+            return call_id, func_name, func_args
+        if hasattr(tool_call, "name") and hasattr(tool_call, "input"):
             call_id = getattr(tool_call, "id", f"call_{id(tool_call)}")
             func_name = sanitize_tool_name(tool_call.name)
-            func_args = tool_call.input  # Already a dict in Anthropic
-        elif isinstance(tool_call, dict):
-            # Support OpenAI "id", Bedrock "toolUseId", or generate one
+            return call_id, func_name, tool_call.input
+        if isinstance(tool_call, dict):
             call_id = (
                 tool_call.get("id")
                 or tool_call.get("toolUseId")
@@ -733,10 +846,15 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 func_info.get("name", "") or tool_call.get("name", "")
             )
             func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})
-        else:
-            return None
+            return call_id, func_name, func_args
+        return None
+
+    def _append_assistant_tool_calls_message(
+        self,
+        parsed_calls: list[tuple[str, str, str | dict[str, Any]]],
+    ) -> None:
+        import json
 
-        # Append assistant message with single tool call
         assistant_message: LLMMessage = {
             "role": "assistant",
             "content": None,
@@ -751,42 +869,54 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                         else json.dumps(func_args),
                     },
                 }
+                for call_id, func_name, func_args in parsed_calls
             ],
         }
-
         self.messages.append(assistant_message)
 
-        # Parse arguments for the single tool call
-        if isinstance(func_args, str):
-            try:
-                args_dict = json.loads(func_args)
-            except json.JSONDecodeError:
-                args_dict = {}
-        else:
-            args_dict = func_args
+    def _execute_single_native_tool_call(
+        self,
+        *,
+        call_id: str,
+        func_name: str,
+        func_args: str | dict[str, Any],
+        available_functions: dict[str, Callable[..., Any]],
+        original_tool: Any | None = None,
+        should_execute: bool = True,
+    ) -> dict[str, Any]:
+        from datetime import datetime
+        import json
 
-        agent_key = getattr(self.agent, "key", "unknown") if self.agent else "unknown"
+        from crewai.events.types.tool_usage_events import (
+            ToolUsageErrorEvent,
+            ToolUsageFinishedEvent,
+            ToolUsageStartedEvent,
+        )
 
-        # Find original tool by matching sanitized name (needed for cache_function and result_as_answer)
+        args_dict, parse_error = parse_tool_call_args(func_args, func_name, call_id, original_tool)
+        if parse_error is not None:
+            return parse_error
 
-        original_tool = None
-        for tool in self.original_tools or []:
-            if sanitize_tool_name(tool.name) == func_name:
-                original_tool = tool
-                break
+        if original_tool is None:
+            for tool in self.original_tools or []:
+                if sanitize_tool_name(tool.name) == func_name:
+                    original_tool = tool
+                    break
 
-        # Check if tool has reached max usage count
         max_usage_reached = False
-        if original_tool:
-            if (
-                hasattr(original_tool, "max_usage_count")
-                and original_tool.max_usage_count is not None
-                and original_tool.current_usage_count >= original_tool.max_usage_count
-            ):
-                max_usage_reached = True
+        if not should_execute and original_tool:
+            max_usage_reached = True
+        elif (
+            should_execute
+            and original_tool
+            and (max_count := getattr(original_tool, "max_usage_count", None))
+            is not None
+            and getattr(original_tool, "current_usage_count", 0) >= max_count
+        ):
+            max_usage_reached = True
 
-        # Check cache before executing
         from_cache = False
+        result: str = "Tool not found"
         input_str = json.dumps(args_dict) if args_dict else ""
         if self.tools_handler and self.tools_handler.cache:
             cached_result = self.tools_handler.cache.read(
@@ -800,7 +930,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 )
                 from_cache = True
 
-        # Emit tool usage started event
+        agent_key = getattr(self.agent, "key", "unknown") if self.agent else "unknown"
         started_at = datetime.now()
         crewai_event_bus.emit(
             self,
@@ -816,14 +946,18 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
 
         track_delegation_if_needed(func_name, args_dict, self.task)
 
-        # Find the structured tool for hook context
         structured_tool: CrewStructuredTool | None = None
-        for structured in self.tools or []:
-            if sanitize_tool_name(structured.name) == func_name:
-                structured_tool = structured
-                break
+        if original_tool is not None:
+            for structured in self.tools or []:
+                if getattr(structured, "_original_tool", None) is original_tool:
+                    structured_tool = structured
+                    break
+        if structured_tool is None:
+            for structured in self.tools or []:
+                if sanitize_tool_name(structured.name) == func_name:
+                    structured_tool = structured
+                    break
 
-        # Execute before_tool_call hooks
         hook_blocked = False
         before_hook_context = ToolCallHookContext(
             tool_name=func_name,
@@ -847,58 +981,48 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                     color="red",
                 )
 
-        # If hook blocked execution, set result and skip tool execution
         if hook_blocked:
             result = f"Tool execution blocked by hook. Tool: {func_name}"
-        # Execute the tool (only if not cached, not at max usage, and not blocked by hook)
-        elif not from_cache and not max_usage_reached:
-            result = "Tool not found"
-            if func_name in available_functions:
-                try:
-                    tool_func = available_functions[func_name]
-                    raw_result = tool_func(**args_dict)
-
-                    # Add to cache after successful execution (before string conversion)
-                    if self.tools_handler and self.tools_handler.cache:
-                        should_cache = True
-                        if (
-                            original_tool
-                            and hasattr(original_tool, "cache_function")
-                            and callable(original_tool.cache_function)
-                        ):
-                            should_cache = original_tool.cache_function(
-                                args_dict, raw_result
-                            )
-                        if should_cache:
-                            self.tools_handler.cache.add(
-                                tool=func_name, input=input_str, output=raw_result
-                            )
-
-                    # Convert to string for message
-                    result = (
-                        str(raw_result)
-                        if not isinstance(raw_result, str)
-                        else raw_result
-                    )
-                except Exception as e:
-                    result = f"Error executing tool: {e}"
-                    if self.task:
-                        self.task.increment_tools_errors()
-                    crewai_event_bus.emit(
-                        self,
-                        event=ToolUsageErrorEvent(
-                            tool_name=func_name,
-                            tool_args=args_dict,
-                            from_agent=self.agent,
-                            from_task=self.task,
-                            agent_key=agent_key,
-                            error=e,
-                        ),
-                    )
-                    error_event_emitted = True
         elif max_usage_reached and original_tool:
-            # Return error message when max usage limit is reached
             result = f"Tool '{func_name}' has reached its usage limit of {original_tool.max_usage_count} times and cannot be used anymore."
+        elif not from_cache and func_name in available_functions:
+            try:
+                raw_result = available_functions[func_name](**args_dict)
+
+                if self.tools_handler and self.tools_handler.cache:
+                    should_cache = True
+                    if (
+                        original_tool
+                        and hasattr(original_tool, "cache_function")
+                        and callable(original_tool.cache_function)
+                    ):
+                        should_cache = original_tool.cache_function(
+                            args_dict, raw_result
+                        )
+                    if should_cache:
+                        self.tools_handler.cache.add(
+                            tool=func_name, input=input_str, output=raw_result
+                        )
+
+                result = (
+                    str(raw_result) if not isinstance(raw_result, str) else raw_result
+                )
+            except Exception as e:
+                result = f"Error executing tool: {e}"
+                if self.task:
+                    self.task.increment_tools_errors()
+                crewai_event_bus.emit(
+                    self,
+                    event=ToolUsageErrorEvent(
+                        tool_name=func_name,
+                        tool_args=args_dict,
+                        from_agent=self.agent,
+                        from_task=self.task,
+                        agent_key=agent_key,
+                        error=e,
+                    ),
+                )
+                error_event_emitted = True
 
         after_hook_context = ToolCallHookContext(
             tool_name=func_name,
@@ -938,7 +1062,23 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 ),
             )
 
-        # Append tool result message
+        return {
+            "call_id": call_id,
+            "func_name": func_name,
+            "result": result,
+            "from_cache": from_cache,
+            "original_tool": original_tool,
+        }
+
+    def _append_tool_result_and_check_finality(
+        self, execution_result: dict[str, Any]
+    ) -> AgentFinish | None:
+        call_id = cast(str, execution_result["call_id"])
+        func_name = cast(str, execution_result["func_name"])
+        result = cast(str, execution_result["result"])
+        from_cache = cast(bool, execution_result["from_cache"])
+        original_tool = execution_result["original_tool"]
+
         tool_message: LLMMessage = {
             "role": "tool",
             "tool_call_id": call_id,
@@ -947,7 +1087,6 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
         }
         self.messages.append(tool_message)
 
-        # Log the tool execution
         if self.agent and self.agent.verbose:
             cache_info = " (from cache)" if from_cache else ""
             self._printer.print(
@@ -960,20 +1099,11 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
             and hasattr(original_tool, "result_as_answer")
             and original_tool.result_as_answer
         ):
-            # Return immediately with tool result as final answer
             return AgentFinish(
                 thought="Tool result is the final answer",
                 output=result,
                 text=result,
             )
-
-        # Inject post-tool reasoning prompt to enforce analysis
-        reasoning_prompt = self._i18n.slice("post_tool_reasoning")
-        reasoning_message: LLMMessage = {
-            "role": "user",
-            "content": reasoning_prompt,
-        }
-        self.messages.append(reasoning_message)
         return None
 
     async def ainvoke(self, inputs: dict[str, Any]) -> dict[str, Any]:
@@ -1133,7 +1263,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                         formatted_answer, tool_result
                     )
 
-                self._invoke_step_callback(formatted_answer)  # type: ignore[arg-type]
+                await self._ainvoke_step_callback(formatted_answer)  # type: ignore[arg-type]
                 self._append_message(formatted_answer.text)  # type: ignore[union-attr]
 
             except OutputParserError as e:
@@ -1186,8 +1316,8 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
         if not self.original_tools:
             return await self._ainvoke_loop_native_no_tools()
 
-        openai_tools, available_functions = convert_tools_to_openai_schema(
-            self.original_tools
+        openai_tools, available_functions, self._tool_name_mapping = (
+            convert_tools_to_openai_schema(self.original_tools)
         )
 
         while True:
@@ -1248,7 +1378,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                         output=answer,
                         text=answer,
                     )
-                    self._invoke_step_callback(formatted_answer)
+                    await self._ainvoke_step_callback(formatted_answer)
                     self._append_message(answer)  # Save final answer to messages
                     self._show_logs(formatted_answer)
                     return formatted_answer
@@ -1260,7 +1390,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                         output=answer,
                         text=output_json,
                     )
-                    self._invoke_step_callback(formatted_answer)
+                    await self._ainvoke_step_callback(formatted_answer)
                     self._append_message(output_json)
                     self._show_logs(formatted_answer)
                     return formatted_answer
@@ -1271,7 +1401,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                     output=str(answer),
                     text=str(answer),
                 )
-                self._invoke_step_callback(formatted_answer)
+                await self._ainvoke_step_callback(formatted_answer)
                 self._append_message(str(answer))  # Save final answer to messages
                 self._show_logs(formatted_answer)
                 return formatted_answer
@@ -1365,13 +1495,28 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
     def _invoke_step_callback(
         self, formatted_answer: AgentAction | AgentFinish
     ) -> None:
-        """Invoke step callback.
+        """Invoke step callback (sync context).
 
         Args:
             formatted_answer: Current agent response.
         """
         if self.step_callback:
-            self.step_callback(formatted_answer)
+            cb_result = self.step_callback(formatted_answer)
+            if inspect.iscoroutine(cb_result):
+                asyncio.run(cb_result)
+
+    async def _ainvoke_step_callback(
+        self, formatted_answer: AgentAction | AgentFinish
+    ) -> None:
+        """Invoke step callback (async context).
+
+        Args:
+            formatted_answer: Current agent response.
+        """
+        if self.step_callback:
+            cb_result = self.step_callback(formatted_answer)
+            if inspect.iscoroutine(cb_result):
+                await cb_result
 
     def _append_message(
         self, text: str, role: Literal["user", "assistant", "system"] = "assistant"

lib/crewai/src/crewai/cli/authentication/__init__.py

@@ -1,5 +1,4 @@
 from crewai.cli.authentication.main import AuthenticationCommand
 
 
-
 __all__ = ["AuthenticationCommand"]

lib/crewai/src/crewai/cli/authentication/main.py

@@ -2,8 +2,8 @@ import time
 from typing import TYPE_CHECKING, Any, TypeVar, cast
 import webbrowser
 
+import httpx
 from pydantic import BaseModel, Field
-import requests
 from rich.console import Console
 
 from crewai.cli.authentication.utils import validate_jwt_token
@@ -98,7 +98,7 @@ class AuthenticationCommand:
             "scope": " ".join(self.oauth2_provider.get_oauth_scopes()),
             "audience": self.oauth2_provider.get_audience(),
         }
-        response = requests.post(
+        response = httpx.post(
             url=self.oauth2_provider.get_authorize_url(),
             data=device_code_payload,
             timeout=20,
@@ -130,7 +130,7 @@ class AuthenticationCommand:
 
         attempts = 0
         while True and attempts < 10:
-            response = requests.post(
+            response = httpx.post(
                 self.oauth2_provider.get_token_url(), data=token_payload, timeout=30
             )
             token_data = response.json()
@@ -149,7 +149,7 @@ class AuthenticationCommand:
                 return
 
             if token_data["error"] not in ("authorization_pending", "slow_down"):
-                raise requests.HTTPError(
+                raise httpx.HTTPError(
                     token_data.get("error_description") or token_data.get("error")
                 )
 

lib/crewai/src/crewai/cli/command.py

@@ -1,5 +1,6 @@
-import requests
-from requests.exceptions import JSONDecodeError
+import json
+
+import httpx
 from rich.console import Console
 
 from crewai.cli.authentication.token import get_auth_token
@@ -30,16 +31,16 @@ class PlusAPIMixin:
             console.print("Run 'crewai login' to sign up/login.", style="bold green")
             raise SystemExit from None
 
-    def _validate_response(self, response: requests.Response) -> None:
+    def _validate_response(self, response: httpx.Response) -> None:
         """
         Handle and display error messages from API responses.
 
         Args:
-            response (requests.Response): The response from the Plus API
+            response (httpx.Response): The response from the Plus API
         """
         try:
             json_response = response.json()
-        except (JSONDecodeError, ValueError):
+        except (json.JSONDecodeError, ValueError):
             console.print(
                 "Failed to parse response from Enterprise API failed. Details:",
                 style="bold red",
@@ -62,7 +63,7 @@ class PlusAPIMixin:
                     )
             raise SystemExit
 
-        if not response.ok:
+        if not response.is_success:
             console.print(
                 "Request to Enterprise API failed. Details:", style="bold red"
             )

lib/crewai/src/crewai/cli/constants.py

@@ -69,7 +69,7 @@ ENV_VARS: dict[str, list[dict[str, Any]]] = {
         },
         {
             "prompt": "Enter your AWS Region Name (press Enter to skip)",
-            "key_name": "AWS_REGION_NAME",
+            "key_name": "AWS_DEFAULT_REGION",
         },
     ],
     "azure": [

lib/crewai/src/crewai/cli/create_crew.py

@@ -143,7 +143,7 @@ def create_folder_structure(
         (folder_path / "src" / folder_name).mkdir(parents=True)
         (folder_path / "src" / folder_name / "tools").mkdir(parents=True)
         (folder_path / "src" / folder_name / "config").mkdir(parents=True)
-        
+
         # Copy AGENTS.md to project root (top-level projects only)
         package_dir = Path(__file__).parent
         agents_md_src = package_dir / "templates" / "AGENTS.md"

lib/crewai/src/crewai/cli/create_flow.py

@@ -1,5 +1,5 @@
-import shutil
 from pathlib import Path
+import shutil
 
 import click
 

lib/crewai/src/crewai/cli/enterprise/main.py

@@ -1,7 +1,7 @@
+import json
 from typing import Any, cast
 
-import requests
-from requests.exceptions import JSONDecodeError, RequestException
+import httpx
 from rich.console import Console
 
 from crewai.cli.authentication.main import Oauth2Settings, ProviderFactory
@@ -47,12 +47,12 @@ class EnterpriseConfigureCommand(BaseCommand):
                 "User-Agent": f"CrewAI-CLI/{get_crewai_version()}",
                 "X-Crewai-Version": get_crewai_version(),
             }
-            response = requests.get(oauth_endpoint, timeout=30, headers=headers)
+            response = httpx.get(oauth_endpoint, timeout=30, headers=headers)
             response.raise_for_status()
 
             try:
                 oauth_config = response.json()
-            except JSONDecodeError as e:
+            except json.JSONDecodeError as e:
                 raise ValueError(f"Invalid JSON response from {oauth_endpoint}") from e
 
             self._validate_oauth_config(oauth_config)
@@ -62,7 +62,7 @@ class EnterpriseConfigureCommand(BaseCommand):
             )
             return cast(dict[str, Any], oauth_config)
 
-        except RequestException as e:
+        except httpx.HTTPError as e:
             raise ValueError(f"Failed to connect to enterprise URL: {e!s}") from e
         except Exception as e:
             raise ValueError(f"Error fetching OAuth2 configuration: {e!s}") from e

lib/crewai/src/crewai/cli/memory_tui.py

@@ -290,13 +290,20 @@ class MemoryTUI(App[None]):
         if self._memory is None:
             panel.update(self._init_error or "No memory loaded.")
             return
+        display_limit = 1000
         info = self._memory.info(path)
         self._last_scope_info = info
-        self._entries = self._memory.list_records(scope=path, limit=200)
+        self._entries = self._memory.list_records(scope=path, limit=display_limit)
         panel.update(_format_scope_info(info))
         panel.border_title = "Detail"
         entry_list = self.query_one("#entry-list", OptionList)
-        entry_list.border_title = f"Entries ({len(self._entries)})"
+        capped = info.record_count > display_limit
+        count_label = (
+            f"Entries (showing {display_limit} of {info.record_count} — display limit)"
+            if capped
+            else f"Entries ({len(self._entries)})"
+        )
+        entry_list.border_title = count_label
         self._populate_entry_list()
 
     def on_option_list_option_highlighted(
@@ -376,6 +383,11 @@ class MemoryTUI(App[None]):
                 return
 
             info_lines: list[str] = []
+            info_lines.append(
+                "[dim italic]Searched the full dataset"
+                + (f" within [bold]{scope}[/]" if scope else "")
+                + " using the recall flow (semantic + recency + importance).[/]\n"
+            )
             if not self._custom_embedder:
                 info_lines.append(
                     "[dim italic]Note: Using default OpenAI embedder. "

lib/crewai/src/crewai/cli/organization/main.py

@@ -1,4 +1,4 @@
-from requests import HTTPError
+from httpx import HTTPStatusError
 from rich.console import Console
 from rich.table import Table
 
@@ -10,11 +10,11 @@ console = Console()
 
 
 class OrganizationCommand(BaseCommand, PlusAPIMixin):
-    def __init__(self):
+    def __init__(self) -> None:
         BaseCommand.__init__(self)
         PlusAPIMixin.__init__(self, telemetry=self._telemetry)
 
-    def list(self):
+    def list(self) -> None:
         try:
             response = self.plus_api_client.get_organizations()
             response.raise_for_status()
@@ -33,7 +33,7 @@ class OrganizationCommand(BaseCommand, PlusAPIMixin):
                 table.add_row(org["name"], org["uuid"])
 
             console.print(table)
-        except HTTPError as e:
+        except HTTPStatusError as e:
             if e.response.status_code == 401:
                 console.print(
                     "You are not logged in to any organization. Use 'crewai login' to login.",
@@ -50,7 +50,7 @@ class OrganizationCommand(BaseCommand, PlusAPIMixin):
             )
             raise SystemExit(1) from e
 
-    def switch(self, org_id):
+    def switch(self, org_id: str) -> None:
         try:
             response = self.plus_api_client.get_organizations()
             response.raise_for_status()
@@ -72,7 +72,7 @@ class OrganizationCommand(BaseCommand, PlusAPIMixin):
                 f"Successfully switched to {org['name']} ({org['uuid']})",
                 style="bold green",
             )
-        except HTTPError as e:
+        except HTTPStatusError as e:
             if e.response.status_code == 401:
                 console.print(
                     "You are not logged in to any organization. Use 'crewai login' to login.",
@@ -87,7 +87,7 @@ class OrganizationCommand(BaseCommand, PlusAPIMixin):
             console.print(f"Failed to switch organization: {e!s}", style="bold red")
             raise SystemExit(1) from e
 
-    def current(self):
+    def current(self) -> None:
         settings = Settings()
         if settings.org_uuid:
             console.print(

lib/crewai/src/crewai/cli/plus_api.py

@@ -3,7 +3,6 @@ from typing import Any
 from urllib.parse import urljoin
 
 import httpx
-import requests
 
 from crewai.cli.config import Settings
 from crewai.cli.constants import DEFAULT_CREWAI_ENTERPRISE_URL
@@ -23,14 +22,15 @@ class PlusAPI:
     EPHEMERAL_TRACING_RESOURCE = "/crewai_plus/api/v1/tracing/ephemeral"
     INTEGRATIONS_RESOURCE = "/crewai_plus/api/v1/integrations"
 
-    def __init__(self, api_key: str) -> None:
+    def __init__(self, api_key: str | None = None) -> None:
         self.api_key = api_key
         self.headers = {
-            "Authorization": f"Bearer {api_key}",
             "Content-Type": "application/json",
             "User-Agent": f"CrewAI-CLI/{get_crewai_version()}",
             "X-Crewai-Version": get_crewai_version(),
         }
+        if api_key:
+            self.headers["Authorization"] = f"Bearer {api_key}"
         settings = Settings()
         if settings.org_uuid:
             self.headers["X-Crewai-Organization-Id"] = settings.org_uuid
@@ -43,16 +43,21 @@ class PlusAPI:
 
     def _make_request(
         self, method: str, endpoint: str, **kwargs: Any
-    ) -> requests.Response:
+    ) -> httpx.Response:
         url = urljoin(self.base_url, endpoint)
-        session = requests.Session()
-        session.trust_env = False
-        return session.request(method, url, headers=self.headers, **kwargs)
+        verify = kwargs.pop("verify", True)
+        with httpx.Client(trust_env=False, verify=verify) as client:
+            return client.request(method, url, headers=self.headers, **kwargs)
 
-    def login_to_tool_repository(self) -> requests.Response:
-        return self._make_request("POST", f"{self.TOOLS_RESOURCE}/login")
+    def login_to_tool_repository(
+        self, user_identifier: str | None = None
+    ) -> httpx.Response:
+        payload = {}
+        if user_identifier:
+            payload["user_identifier"] = user_identifier
+        return self._make_request("POST", f"{self.TOOLS_RESOURCE}/login", json=payload)
 
-    def get_tool(self, handle: str) -> requests.Response:
+    def get_tool(self, handle: str) -> httpx.Response:
         return self._make_request("GET", f"{self.TOOLS_RESOURCE}/{handle}")
 
     async def get_agent(self, handle: str) -> httpx.Response:
@@ -68,7 +73,7 @@ class PlusAPI:
         description: str | None,
         encoded_file: str,
         available_exports: list[dict[str, Any]] | None = None,
-    ) -> requests.Response:
+    ) -> httpx.Response:
         params = {
             "handle": handle,
             "public": is_public,
@@ -79,54 +84,52 @@ class PlusAPI:
         }
         return self._make_request("POST", f"{self.TOOLS_RESOURCE}", json=params)
 
-    def deploy_by_name(self, project_name: str) -> requests.Response:
+    def deploy_by_name(self, project_name: str) -> httpx.Response:
         return self._make_request(
             "POST", f"{self.CREWS_RESOURCE}/by-name/{project_name}/deploy"
         )
 
-    def deploy_by_uuid(self, uuid: str) -> requests.Response:
+    def deploy_by_uuid(self, uuid: str) -> httpx.Response:
         return self._make_request("POST", f"{self.CREWS_RESOURCE}/{uuid}/deploy")
 
-    def crew_status_by_name(self, project_name: str) -> requests.Response:
+    def crew_status_by_name(self, project_name: str) -> httpx.Response:
         return self._make_request(
             "GET", f"{self.CREWS_RESOURCE}/by-name/{project_name}/status"
         )
 
-    def crew_status_by_uuid(self, uuid: str) -> requests.Response:
+    def crew_status_by_uuid(self, uuid: str) -> httpx.Response:
         return self._make_request("GET", f"{self.CREWS_RESOURCE}/{uuid}/status")
 
     def crew_by_name(
         self, project_name: str, log_type: str = "deployment"
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "GET", f"{self.CREWS_RESOURCE}/by-name/{project_name}/logs/{log_type}"
         )
 
-    def crew_by_uuid(
-        self, uuid: str, log_type: str = "deployment"
-    ) -> requests.Response:
+    def crew_by_uuid(self, uuid: str, log_type: str = "deployment") -> httpx.Response:
         return self._make_request(
             "GET", f"{self.CREWS_RESOURCE}/{uuid}/logs/{log_type}"
         )
 
-    def delete_crew_by_name(self, project_name: str) -> requests.Response:
+    def delete_crew_by_name(self, project_name: str) -> httpx.Response:
         return self._make_request(
             "DELETE", f"{self.CREWS_RESOURCE}/by-name/{project_name}"
         )
 
-    def delete_crew_by_uuid(self, uuid: str) -> requests.Response:
+    def delete_crew_by_uuid(self, uuid: str) -> httpx.Response:
         return self._make_request("DELETE", f"{self.CREWS_RESOURCE}/{uuid}")
 
-    def list_crews(self) -> requests.Response:
+    def list_crews(self) -> httpx.Response:
         return self._make_request("GET", self.CREWS_RESOURCE)
 
-    def create_crew(self, payload: dict[str, Any]) -> requests.Response:
+    def create_crew(self, payload: dict[str, Any]) -> httpx.Response:
         return self._make_request("POST", self.CREWS_RESOURCE, json=payload)
 
-    def get_organizations(self) -> requests.Response:
+    def get_organizations(self) -> httpx.Response:
         return self._make_request("GET", self.ORGANIZATIONS_RESOURCE)
 
-    def initialize_trace_batch(self, payload: dict[str, Any]) -> requests.Response:
+    def initialize_trace_batch(self, payload: dict[str, Any]) -> httpx.Response:
         return self._make_request(
             "POST",
             f"{self.TRACING_RESOURCE}/batches",
@@ -136,7 +139,7 @@ class PlusAPI:
 
     def initialize_ephemeral_trace_batch(
         self, payload: dict[str, Any]
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "POST",
             f"{self.EPHEMERAL_TRACING_RESOURCE}/batches",
@@ -145,7 +148,7 @@ class PlusAPI:
 
     def send_trace_events(
         self, trace_batch_id: str, payload: dict[str, Any]
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "POST",
             f"{self.TRACING_RESOURCE}/batches/{trace_batch_id}/events",
@@ -155,7 +158,7 @@ class PlusAPI:
 
     def send_ephemeral_trace_events(
         self, trace_batch_id: str, payload: dict[str, Any]
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "POST",
             f"{self.EPHEMERAL_TRACING_RESOURCE}/batches/{trace_batch_id}/events",
@@ -165,7 +168,7 @@ class PlusAPI:
 
     def finalize_trace_batch(
         self, trace_batch_id: str, payload: dict[str, Any]
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "PATCH",
             f"{self.TRACING_RESOURCE}/batches/{trace_batch_id}/finalize",
@@ -175,7 +178,7 @@ class PlusAPI:
 
     def finalize_ephemeral_trace_batch(
         self, trace_batch_id: str, payload: dict[str, Any]
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "PATCH",
             f"{self.EPHEMERAL_TRACING_RESOURCE}/batches/{trace_batch_id}/finalize",
@@ -185,7 +188,7 @@ class PlusAPI:
 
     def mark_trace_batch_as_failed(
         self, trace_batch_id: str, error_message: str
-    ) -> requests.Response:
+    ) -> httpx.Response:
         return self._make_request(
             "PATCH",
             f"{self.TRACING_RESOURCE}/batches/{trace_batch_id}",
@@ -193,13 +196,20 @@ class PlusAPI:
             timeout=30,
         )
 
-    def get_triggers(self) -> requests.Response:
+    def get_mcp_configs(self, slugs: list[str]) -> httpx.Response:
+        """Get MCP server configurations for the given slugs."""
+        return self._make_request(
+            "GET",
+            f"{self.INTEGRATIONS_RESOURCE}/mcp_configs",
+            params={"slugs": ",".join(slugs)},
+            timeout=30,
+        )
+
+    def get_triggers(self) -> httpx.Response:
         """Get all available triggers from integrations."""
         return self._make_request("GET", f"{self.INTEGRATIONS_RESOURCE}/apps")
 
-    def get_trigger_payload(
-        self, app_slug: str, trigger_slug: str
-    ) -> requests.Response:
+    def get_trigger_payload(self, app_slug: str, trigger_slug: str) -> httpx.Response:
         """Get sample payload for a specific trigger."""
         return self._make_request(
             "GET", f"{self.INTEGRATIONS_RESOURCE}/{app_slug}/{trigger_slug}/payload"

lib/crewai/src/crewai/cli/provider.py

@@ -8,7 +8,7 @@ from typing import Any
 
 import certifi
 import click
-import requests
+import httpx
 
 from crewai.cli.constants import JSON_URL, MODELS, PROVIDERS
 
@@ -165,20 +165,20 @@ def fetch_provider_data(cache_file: Path) -> dict[str, Any] | None:
     ssl_config = os.environ["SSL_CERT_FILE"] = certifi.where()
 
     try:
-        response = requests.get(JSON_URL, stream=True, timeout=60, verify=ssl_config)
-        response.raise_for_status()
-        data = download_data(response)
-        with open(cache_file, "w") as f:
-            json.dump(data, f)
-        return data
-    except requests.RequestException as e:
+        with httpx.stream("GET", JSON_URL, timeout=60, verify=ssl_config) as response:
+            response.raise_for_status()
+            data = download_data(response)
+            with open(cache_file, "w") as f:
+                json.dump(data, f)
+            return data
+    except httpx.HTTPError as e:
         click.secho(f"Error fetching provider data: {e}", fg="red")
     except json.JSONDecodeError:
         click.secho("Error parsing provider data. Invalid JSON format.", fg="red")
     return None
 
 
-def download_data(response: requests.Response) -> dict[str, Any]:
+def download_data(response: httpx.Response) -> dict[str, Any]:
     """Downloads data from a given HTTP response and returns the JSON content.
 
     Args:
@@ -194,7 +194,7 @@ def download_data(response: requests.Response) -> dict[str, Any]:
     with click.progressbar(
         length=total_size, label="Downloading", show_pos=True
     ) as bar:
-        for chunk in response.iter_content(block_size):
+        for chunk in response.iter_bytes(block_size):
             if chunk:
                 data_chunks.append(chunk)
                 bar.update(len(chunk))

lib/crewai/src/crewai/cli/templates/crew/pyproject.toml

@@ -5,7 +5,7 @@ description = "{{name}} using crewAI"
 authors = [{ name = "Your Name", email = "you@example.com" }]
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]==1.9.3"
+    "crewai[tools]==1.10.1"
 ]
 
 [project.scripts]

lib/crewai/src/crewai/cli/templates/flow/pyproject.toml

@@ -5,7 +5,7 @@ description = "{{name}} using crewAI"
 authors = [{ name = "Your Name", email = "you@example.com" }]
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]==1.9.3"
+    "crewai[tools]==1.10.1"
 ]
 
 [project.scripts]

lib/crewai/src/crewai/cli/templates/tool/pyproject.toml

@@ -5,7 +5,7 @@ description = "Power up your crews with {{folder_name}}"
 readme = "README.md"
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]>=0.203.1"
+    "crewai[tools]==1.10.1"
 ]
 
 [tool.crewai]

lib/crewai/src/crewai/cli/tools/main.py

@@ -23,6 +23,7 @@ from crewai.cli.utils import (
     tree_copy,
     tree_find_and_replace,
 )
+from crewai.events.listeners.tracing.utils import get_user_id
 
 
 console = Console()
@@ -169,7 +170,9 @@ class ToolCommand(BaseCommand, PlusAPIMixin):
         console.print(f"Successfully installed {handle}", style="bold green")
 
     def login(self) -> None:
-        login_response = self.plus_api_client.login_to_tool_repository()
+        login_response = self.plus_api_client.login_to_tool_repository(
+            user_identifier=get_user_id()
+        )
 
         if login_response.status_code != 200:
             console.print(

lib/crewai/src/crewai/crews/__init__.py

@@ -1,5 +1,4 @@
 from crewai.crews.crew_output import CrewOutput
 
 
-
 __all__ = ["CrewOutput"]

lib/crewai/src/crewai/events/__init__.py

@@ -63,6 +63,7 @@ from crewai.events.types.logging_events import (
     AgentLogsStartedEvent,
 )
 from crewai.events.types.mcp_events import (
+    MCPConfigFetchFailedEvent,
     MCPConnectionCompletedEvent,
     MCPConnectionFailedEvent,
     MCPConnectionStartedEvent,
@@ -165,6 +166,7 @@ __all__ = [
     "LiteAgentExecutionCompletedEvent",
     "LiteAgentExecutionErrorEvent",
     "LiteAgentExecutionStartedEvent",
+    "MCPConfigFetchFailedEvent",
     "MCPConnectionCompletedEvent",
     "MCPConnectionFailedEvent",
     "MCPConnectionStartedEvent",

lib/crewai/src/crewai/events/base_event_listener.py

@@ -23,4 +23,3 @@ class BaseEventListener(ABC):
         Args:
             crewai_event_bus: The event bus to register listeners on.
         """
-        pass

lib/crewai/src/crewai/events/event_listener.py

@@ -68,6 +68,7 @@ from crewai.events.types.logging_events import (
     AgentLogsStartedEvent,
 )
 from crewai.events.types.mcp_events import (
+    MCPConfigFetchFailedEvent,
     MCPConnectionCompletedEvent,
     MCPConnectionFailedEvent,
     MCPConnectionStartedEvent,
@@ -665,6 +666,16 @@ class EventListener(BaseEventListener):
                 event.error_type,
             )
 
+        @crewai_event_bus.on(MCPConfigFetchFailedEvent)
+        def on_mcp_config_fetch_failed(
+            _: Any, event: MCPConfigFetchFailedEvent
+        ) -> None:
+            self.formatter.handle_mcp_config_fetch_failed(
+                event.slug,
+                event.error,
+                event.error_type,
+            )
+
         @crewai_event_bus.on(MCPToolExecutionStartedEvent)
         def on_mcp_tool_execution_started(
             _: Any, event: MCPToolExecutionStartedEvent

lib/crewai/src/crewai/events/event_types.py

@@ -67,6 +67,7 @@ from crewai.events.types.llm_guardrail_events import (
     LLMGuardrailStartedEvent,
 )
 from crewai.events.types.mcp_events import (
+    MCPConfigFetchFailedEvent,
     MCPConnectionCompletedEvent,
     MCPConnectionFailedEvent,
     MCPConnectionStartedEvent,
@@ -181,4 +182,5 @@ EventTypes = (
     | MCPToolExecutionStartedEvent
     | MCPToolExecutionCompletedEvent
     | MCPToolExecutionFailedEvent
+    | MCPConfigFetchFailedEvent
 )

lib/crewai/src/crewai/events/listeners/tracing/trace_batch_manager.py

@@ -15,6 +15,7 @@ from crewai.cli.plus_api import PlusAPI
 from crewai.cli.version import get_crewai_version
 from crewai.events.listeners.tracing.types import TraceEvent
 from crewai.events.listeners.tracing.utils import (
+    get_user_id,
     is_tracing_enabled_in_context,
     should_auto_collect_first_time_traces,
 )
@@ -67,7 +68,7 @@ class TraceBatchManager:
                 api_key=get_auth_token(),
             )
         except AuthError:
-            self.plus_api = PlusAPI(api_key="")
+            self.plus_api = PlusAPI()
         self.ephemeral_trace_url = None
 
     def initialize_batch(
@@ -120,7 +121,6 @@ class TraceBatchManager:
             payload = {
                 "trace_id": self.current_batch.batch_id,
                 "execution_type": execution_metadata.get("execution_type", "crew"),
-                "user_identifier": execution_metadata.get("user_context", None),
                 "execution_context": {
                     "crew_fingerprint": execution_metadata.get("crew_fingerprint"),
                     "crew_name": execution_metadata.get("crew_name", None),
@@ -140,6 +140,7 @@ class TraceBatchManager:
             }
             if use_ephemeral:
                 payload["ephemeral_trace_id"] = self.current_batch.batch_id
+                payload["user_identifier"] = get_user_id()
 
             response = (
                 self.plus_api.initialize_ephemeral_trace_batch(payload)

lib/crewai/src/crewai/events/types/llm_events.py

@@ -86,3 +86,11 @@ class LLMStreamChunkEvent(LLMEventBase):
     tool_call: ToolCall | None = None
     call_type: LLMCallType | None = None
     response_id: str | None = None
+
+
+class LLMThinkingChunkEvent(LLMEventBase):
+    """Event emitted when a thinking/reasoning chunk is received from a thinking model"""
+
+    type: str = "llm_thinking_chunk"
+    chunk: str
+    response_id: str | None = None

lib/crewai/src/crewai/events/types/mcp_events.py

@@ -83,3 +83,16 @@ class MCPToolExecutionFailedEvent(MCPEvent):
     error_type: str | None = None  # "timeout", "validation", "server_error", etc.
     started_at: datetime | None = None
     failed_at: datetime | None = None
+
+
+class MCPConfigFetchFailedEvent(BaseEvent):
+    """Event emitted when fetching an AMP MCP server config fails.
+
+    This covers cases where the slug is not connected, the API call
+    failed, or native MCP resolution failed after config was fetched.
+    """
+
+    type: str = "mcp_config_fetch_failed"
+    slug: str
+    error: str
+    error_type: str | None = None  # "not_connected", "api_error", "connection_failed"

lib/crewai/src/crewai/events/types/reasoning_events.py

@@ -9,7 +9,7 @@ class ReasoningEvent(BaseEvent):
     type: str
     attempt: int = 1
     agent_role: str
-    task_id: str
+    task_id: str | None = None
     task_name: str | None = None
     from_task: Any | None = None
     agent_id: str | None = None

lib/crewai/src/crewai/events/utils/console_formatter.py

@@ -1512,6 +1512,34 @@ To enable tracing, do any one of these:
         self.print(panel)
         self.print()
 
+    def handle_mcp_config_fetch_failed(
+        self,
+        slug: str,
+        error: str = "",
+        error_type: str | None = None,
+    ) -> None:
+        """Handle MCP config fetch failed event (AMP resolution failures)."""
+        if not self.verbose:
+            return
+
+        content = Text()
+        content.append("MCP Config Fetch Failed\n\n", style="red bold")
+        content.append("Server: ", style="white")
+        content.append(f"{slug}\n", style="red")
+
+        if error_type:
+            content.append("Error Type: ", style="white")
+            content.append(f"{error_type}\n", style="red")
+
+        if error:
+            content.append("\nError: ", style="white bold")
+            error_preview = error[:500] + "..." if len(error) > 500 else error
+            content.append(f"{error_preview}\n", style="red")
+
+        panel = self.create_panel(content, "❌ MCP Config Failed", "red")
+        self.print(panel)
+        self.print()
+
     def handle_mcp_tool_execution_started(
         self,
         server_name: str,

lib/crewai/src/crewai/experimental/agent_executor.py

@@ -1,7 +1,10 @@
 from __future__ import annotations
 
+import asyncio
 from collections.abc import Callable, Coroutine
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from datetime import datetime
+import inspect
 import json
 import threading
 from typing import TYPE_CHECKING, Any, Literal, cast
@@ -49,6 +52,8 @@ from crewai.hooks.types import (
     BeforeLLMCallHookCallable,
     BeforeLLMCallHookType,
 )
+from crewai.tools.base_tool import BaseTool
+from crewai.tools.structured_tool import CrewStructuredTool
 from crewai.utilities.agent_utils import (
     convert_tools_to_openai_schema,
     enforce_rpm_limit,
@@ -63,11 +68,13 @@ from crewai.utilities.agent_utils import (
     has_reached_max_iterations,
     is_context_length_exceeded,
     is_inside_event_loop,
+    parse_tool_call_args,
     process_llm_response,
     track_delegation_if_needed,
 )
 from crewai.utilities.constants import TRAINING_DATA_FILE
 from crewai.utilities.i18n import I18N, get_i18n
+from crewai.utilities.planning_types import PlanStep, TodoItem, TodoList
 from crewai.utilities.printer import Printer
 from crewai.utilities.string_utils import sanitize_tool_name
 from crewai.utilities.tool_utils import execute_tool_and_check_finality
@@ -81,8 +88,6 @@ if TYPE_CHECKING:
     from crewai.crew import Crew
     from crewai.llms.base_llm import BaseLLM
     from crewai.task import Task
-    from crewai.tools.base_tool import BaseTool
-    from crewai.tools.structured_tool import CrewStructuredTool
     from crewai.tools.tool_types import ToolResult
     from crewai.utilities.prompts import StandardPromptResult, SystemPromptResult
 
@@ -101,6 +106,13 @@ class AgentReActState(BaseModel):
     ask_for_human_input: bool = Field(default=False)
     use_native_tools: bool = Field(default=False)
     pending_tool_calls: list[Any] = Field(default_factory=list)
+    plan: str | None = Field(default=None, description="Generated execution plan")
+    plan_ready: bool = Field(
+        default=False, description="Whether agent is ready to execute"
+    )
+    todos: TodoList = Field(
+        default_factory=TodoList, description="Todo list for tracking plan execution"
+    )
 
 
 class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
@@ -298,6 +310,7 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
             super().__init__(
                 suppress_flow_events=True,
                 tracing=current_tracing if current_tracing else None,
+                max_method_calls=self.max_iter * 10,
             )
             self._flow_initialized = True
 
@@ -317,7 +330,7 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
     def _setup_native_tools(self) -> None:
         """Convert tools to OpenAI schema format for native function calling."""
         if self.original_tools:
-            self._openai_tools, self._available_functions = (
+            self._openai_tools, self._available_functions, self._tool_name_mapping = (
                 convert_tools_to_openai_schema(self.original_tools)
             )
 
@@ -389,6 +402,67 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
         self._state.iterations = value
 
     @start()
+    def generate_plan(self) -> None:
+        """Generate execution plan if planning is enabled.
+
+        This is the entry point for the agent execution flow. If planning is
+        enabled on the agent, it generates a plan before execution begins.
+        The plan is stored in state and todos are created from the steps.
+        """
+        if not getattr(self.agent, "planning_enabled", False):
+            return
+
+        try:
+            from crewai.utilities.reasoning_handler import AgentReasoning
+
+            if self.task:
+                planning_handler = AgentReasoning(agent=self.agent, task=self.task)
+            else:
+                # For kickoff() path - use input text directly, no Task needed
+                input_text = getattr(self, "_kickoff_input", "")
+                planning_handler = AgentReasoning(
+                    agent=self.agent,
+                    description=input_text or "Complete the requested task",
+                    expected_output="Complete the task successfully",
+                )
+
+            output = planning_handler.handle_agent_reasoning()
+
+            self.state.plan = output.plan.plan
+            self.state.plan_ready = output.plan.ready
+
+            if self.state.plan_ready and output.plan.steps:
+                self._create_todos_from_plan(output.plan.steps)
+
+            # Backward compatibility: append plan to task description
+            # This can be removed in Phase 2 when plan execution is implemented
+            if self.task and self.state.plan:
+                self.task.description += f"\n\nPlanning:\n{self.state.plan}"
+
+        except Exception as e:
+            if hasattr(self.agent, "_logger"):
+                self.agent._logger.log("error", f"Error during planning: {e!s}")
+
+    def _create_todos_from_plan(self, steps: list[PlanStep]) -> None:
+        """Convert plan steps into trackable todo items.
+
+        Args:
+            steps: List of PlanStep objects from the reasoning handler.
+        """
+        todos: list[TodoItem] = []
+        for step in steps:
+            todo = TodoItem(
+                step_number=step.step_number,
+                description=step.description,
+                tool_to_use=step.tool_to_use,
+                depends_on=step.depends_on,
+                status="pending",
+            )
+            todos.append(todo)
+
+        self.state.todos = TodoList(items=todos)
+
+    @listen(generate_plan)
     def initialize_reasoning(self) -> Literal["initialized"]:
         """Initialize the reasoning flow and emit agent start logs."""
         self._show_start_logs()
@@ -399,7 +473,7 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
                 self._setup_native_tools()
         return "initialized"
 
-    @listen("force_final_answer")
+    @listen("max_iterations_exceeded")
     def force_final_answer(self) -> Literal["agent_finished"]:
         """Force agent to provide final answer when max iterations exceeded."""
         formatted_answer = handle_max_iterations_exceeded(
@@ -590,21 +664,19 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
     def execute_tool_action(self) -> Literal["tool_completed", "tool_result_is_final"]:
         """Execute the tool action and handle the result."""
 
+        action = cast(AgentAction, self.state.current_answer)
+
+        fingerprint_context = {}
+        if (
+            self.agent
+            and hasattr(self.agent, "security_config")
+            and hasattr(self.agent.security_config, "fingerprint")
+        ):
+            fingerprint_context = {
+                "agent_fingerprint": str(self.agent.security_config.fingerprint)
+            }
+
         try:
-            action = cast(AgentAction, self.state.current_answer)
-
-            # Extract fingerprint context for tool execution
-            fingerprint_context = {}
-            if (
-                self.agent
-                and hasattr(self.agent, "security_config")
-                and hasattr(self.agent.security_config, "fingerprint")
-            ):
-                fingerprint_context = {
-                    "agent_fingerprint": str(self.agent.security_config.fingerprint)
-                }
-
-            # Execute the tool
             tool_result = execute_tool_and_check_finality(
                 agent_action=action,
                 fingerprint_context=fingerprint_context,
@@ -618,24 +690,19 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
                 function_calling_llm=self.function_calling_llm,
                 crew=self.crew,
             )
+        except Exception as e:
+            if self.agent and self.agent.verbose:
+                self._printer.print(
+                    content=f"Error in tool execution: {e}", color="red"
+                )
+            if self.task:
+                self.task.increment_tools_errors()
 
-            # Handle agent action and append observation to messages
-            result = self._handle_agent_action(action, tool_result)
-            self.state.current_answer = result
+            error_observation = f"\nObservation: Error executing tool: {e}"
+            action.text += error_observation
+            action.result = str(e)
+            self._append_message_to_state(action.text)
 
-            # Invoke step callback if configured
-            self._invoke_step_callback(result)
-
-            # Append result message to conversation state
-            if hasattr(result, "text"):
-                self._append_message_to_state(result.text)
-
-            # Check if tool result became a final answer (result_as_answer flag)
-            if isinstance(result, AgentFinish):
-                self.state.is_finished = True
-                return "tool_result_is_final"
-
-            # Inject post-tool reasoning prompt to enforce analysis
             reasoning_prompt = self._i18n.slice("post_tool_reasoning")
             reasoning_message: LLMMessage = {
                 "role": "user",
@@ -645,12 +712,26 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
 
             return "tool_completed"
 
-        except Exception as e:
-            error_text = Text()
-            error_text.append("❌ Error in tool execution: ", style="red bold")
-            error_text.append(str(e), style="red")
-            self._console.print(error_text)
-            raise
+        result = self._handle_agent_action(action, tool_result)
+        self.state.current_answer = result
+
+        self._invoke_step_callback(result)
+
+        if hasattr(result, "text"):
+            self._append_message_to_state(result.text)
+
+        if isinstance(result, AgentFinish):
+            self.state.is_finished = True
+            return "tool_result_is_final"
+
+        reasoning_prompt = self._i18n.slice("post_tool_reasoning")
+        reasoning_message_post: LLMMessage = {
+            "role": "user",
+            "content": reasoning_prompt,
+        }
+        self.state.messages.append(reasoning_message_post)
+
+        return "tool_completed"
 
     @listen("native_tool_calls")
     def execute_native_tool(
@@ -668,9 +749,12 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
         if not self.state.pending_tool_calls:
             return "native_tool_completed"
 
+        pending_tool_calls = list(self.state.pending_tool_calls)
+        self.state.pending_tool_calls.clear()
+
         # Group all tool calls into a single assistant message
         tool_calls_to_report = []
-        for tool_call in self.state.pending_tool_calls:
+        for tool_call in pending_tool_calls:
             info = extract_tool_call_info(tool_call)
             if not info:
                 continue
@@ -695,202 +779,99 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
                 "content": None,
                 "tool_calls": tool_calls_to_report,
             }
-            if all(
-                type(tc).__qualname__ == "Part" for tc in self.state.pending_tool_calls
-            ):
-                assistant_message["raw_tool_call_parts"] = list(
-                    self.state.pending_tool_calls
-                )
+            if all(type(tc).__qualname__ == "Part" for tc in pending_tool_calls):
+                assistant_message["raw_tool_call_parts"] = list(pending_tool_calls)
             self.state.messages.append(assistant_message)
 
-        # Now execute each tool
-        while self.state.pending_tool_calls:
-            tool_call = self.state.pending_tool_calls.pop(0)
-            info = extract_tool_call_info(tool_call)
-            if not info:
-                continue
+        runnable_tool_calls = [
+            tool_call
+            for tool_call in pending_tool_calls
+            if extract_tool_call_info(tool_call) is not None
+        ]
+        should_parallelize = self._should_parallelize_native_tool_calls(
+            runnable_tool_calls
+        )
 
-            call_id, func_name, func_args = info
-
-            # Parse arguments
-            if isinstance(func_args, str):
-                try:
-                    args_dict = json.loads(func_args)
-                except json.JSONDecodeError:
-                    args_dict = {}
-            else:
-                args_dict = func_args
-
-            # Get agent_key for event tracking
-            agent_key = (
-                getattr(self.agent, "key", "unknown") if self.agent else "unknown"
-            )
-
-            # Find original tool by matching sanitized name (needed for cache_function and result_as_answer)
-            original_tool = None
-            for tool in self.original_tools or []:
-                if sanitize_tool_name(tool.name) == func_name:
-                    original_tool = tool
-                    break
-
-            # Check if tool has reached max usage count
-            max_usage_reached = False
-            if (
-                original_tool
-                and original_tool.max_usage_count is not None
-                and original_tool.current_usage_count >= original_tool.max_usage_count
-            ):
-                max_usage_reached = True
-
-            # Check cache before executing
-            from_cache = False
-            input_str = json.dumps(args_dict) if args_dict else ""
-            if self.tools_handler and self.tools_handler.cache:
-                cached_result = self.tools_handler.cache.read(
-                    tool=func_name, input=input_str
+        execution_results: list[dict[str, Any]] = []
+        if should_parallelize:
+            max_workers = min(8, len(runnable_tool_calls))
+            with ThreadPoolExecutor(max_workers=max_workers) as pool:
+                future_to_idx = {
+                    pool.submit(self._execute_single_native_tool_call, tool_call): idx
+                    for idx, tool_call in enumerate(runnable_tool_calls)
+                }
+                ordered_results: list[dict[str, Any] | None] = [None] * len(
+                    runnable_tool_calls
                 )
-                if cached_result is not None:
-                    result = (
-                        str(cached_result)
-                        if not isinstance(cached_result, str)
-                        else cached_result
-                    )
-                    from_cache = True
-
-            # Emit tool usage started event
-            started_at = datetime.now()
-            crewai_event_bus.emit(
-                self,
-                event=ToolUsageStartedEvent(
-                    tool_name=func_name,
-                    tool_args=args_dict,
-                    from_agent=self.agent,
-                    from_task=self.task,
-                    agent_key=agent_key,
-                ),
-            )
-            error_event_emitted = False
-
-            track_delegation_if_needed(func_name, args_dict, self.task)
-
-            structured_tool: CrewStructuredTool | None = None
-            for structured in self.tools or []:
-                if sanitize_tool_name(structured.name) == func_name:
-                    structured_tool = structured
-                    break
-
-            hook_blocked = False
-            before_hook_context = ToolCallHookContext(
-                tool_name=func_name,
-                tool_input=args_dict,
-                tool=structured_tool,  # type: ignore[arg-type]
-                agent=self.agent,
-                task=self.task,
-                crew=self.crew,
-            )
-            before_hooks = get_before_tool_call_hooks()
-            try:
-                for hook in before_hooks:
-                    hook_result = hook(before_hook_context)
-                    if hook_result is False:
-                        hook_blocked = True
-                        break
-            except Exception as hook_error:
-                if self.agent.verbose:
-                    self._printer.print(
-                        content=f"Error in before_tool_call hook: {hook_error}",
-                        color="red",
-                    )
-
-            if hook_blocked:
-                result = f"Tool execution blocked by hook. Tool: {func_name}"
-            elif not from_cache and not max_usage_reached:
-                result = "Tool not found"
-                if func_name in self._available_functions:
+                for future in as_completed(future_to_idx):
+                    idx = future_to_idx[future]
                     try:
-                        tool_func = self._available_functions[func_name]
-                        raw_result = tool_func(**args_dict)
-
-                        # Add to cache after successful execution (before string conversion)
-                        if self.tools_handler and self.tools_handler.cache:
-                            should_cache = True
-                            if original_tool:
-                                should_cache = original_tool.cache_function(
-                                    args_dict, raw_result
-                                )
-                            if should_cache:
-                                self.tools_handler.cache.add(
-                                    tool=func_name, input=input_str, output=raw_result
-                                )
-
-                        # Convert to string for message
-                        result = (
-                            str(raw_result)
-                            if not isinstance(raw_result, str)
-                            else raw_result
-                        )
+                        ordered_results[idx] = future.result()
                     except Exception as e:
-                        result = f"Error executing tool: {e}"
-                        if self.task:
-                            self.task.increment_tools_errors()
-                        # Emit tool usage error event
-                        crewai_event_bus.emit(
-                            self,
-                            event=ToolUsageErrorEvent(
-                                tool_name=func_name,
-                                tool_args=args_dict,
-                                from_agent=self.agent,
-                                from_task=self.task,
-                                agent_key=agent_key,
-                                error=e,
-                            ),
-                        )
-                        error_event_emitted = True
-            elif max_usage_reached and original_tool:
-                # Return error message when max usage limit is reached
-                result = f"Tool '{func_name}' has reached its usage limit of {original_tool.max_usage_count} times and cannot be used anymore."
+                        tool_call = runnable_tool_calls[idx]
+                        info = extract_tool_call_info(tool_call)
+                        call_id = info[0] if info else "unknown"
+                        func_name = info[1] if info else "unknown"
+                        ordered_results[idx] = {
+                            "call_id": call_id,
+                            "func_name": func_name,
+                            "result": f"Error executing tool: {e}",
+                            "from_cache": False,
+                            "original_tool": None,
+                        }
+                execution_results = [
+                    result for result in ordered_results if result is not None
+                ]
+        else:
+            # Execute sequentially so result_as_answer tools can short-circuit
+            # immediately without running remaining calls.
+            for tool_call in runnable_tool_calls:
+                execution_result = self._execute_single_native_tool_call(tool_call)
+                call_id = cast(str, execution_result["call_id"])
+                func_name = cast(str, execution_result["func_name"])
+                result = cast(str, execution_result["result"])
+                from_cache = cast(bool, execution_result["from_cache"])
+                original_tool = execution_result["original_tool"]
 
-            # Execute after_tool_call hooks (even if blocked, to allow logging/monitoring)
-            after_hook_context = ToolCallHookContext(
-                tool_name=func_name,
-                tool_input=args_dict,
-                tool=structured_tool,  # type: ignore[arg-type]
-                agent=self.agent,
-                task=self.task,
-                crew=self.crew,
-                tool_result=result,
-            )
-            after_hooks = get_after_tool_call_hooks()
-            try:
-                for after_hook in after_hooks:
-                    after_hook_result = after_hook(after_hook_context)
-                    if after_hook_result is not None:
-                        result = after_hook_result
-                        after_hook_context.tool_result = result
-            except Exception as hook_error:
-                if self.agent.verbose:
+                tool_message: LLMMessage = {
+                    "role": "tool",
+                    "tool_call_id": call_id,
+                    "name": func_name,
+                    "content": result,
+                }
+                self.state.messages.append(tool_message)
+
+                # Log the tool execution
+                if self.agent and self.agent.verbose:
+                    cache_info = " (from cache)" if from_cache else ""
                     self._printer.print(
-                        content=f"Error in after_tool_call hook: {hook_error}",
-                        color="red",
+                        content=f"Tool {func_name} executed with result{cache_info}: {result[:200]}...",
+                        color="green",
                     )
 
-            if not error_event_emitted:
-                crewai_event_bus.emit(
-                    self,
-                    event=ToolUsageFinishedEvent(
+                if (
+                    original_tool
+                    and hasattr(original_tool, "result_as_answer")
+                    and original_tool.result_as_answer
+                ):
+                    self.state.current_answer = AgentFinish(
+                        thought="Tool result is the final answer",
                         output=result,
-                        tool_name=func_name,
-                        tool_args=args_dict,
-                        from_agent=self.agent,
-                        from_task=self.task,
-                        agent_key=agent_key,
-                        started_at=started_at,
-                        finished_at=datetime.now(),
-                    ),
-                )
+                        text=result,
+                    )
+                    self.state.is_finished = True
+                    return "tool_result_is_final"
 
-            # Append tool result message
-            tool_message: LLMMessage = {
+            return "native_tool_completed"
+
+        for execution_result in execution_results:
+            call_id = cast(str, execution_result["call_id"])
+            func_name = cast(str, execution_result["func_name"])
+            result = cast(str, execution_result["result"])
+            from_cache = cast(bool, execution_result["from_cache"])
+            original_tool = execution_result["original_tool"]
+
+            tool_message = {
                 "role": "tool",
                 "tool_call_id": call_id,
                 "name": func_name,
@@ -922,6 +903,249 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
 
         return "native_tool_completed"
 
+    def _should_parallelize_native_tool_calls(self, tool_calls: list[Any]) -> bool:
+        """Determine if native tool calls are safe to run in parallel."""
+        if len(tool_calls) <= 1:
+            return False
+
+        for tool_call in tool_calls:
+            info = extract_tool_call_info(tool_call)
+            if not info:
+                continue
+            _, func_name, _ = info
+
+            mapping = getattr(self, "_tool_name_mapping", None)
+            original_tool: BaseTool | None = None
+            if mapping and func_name in mapping:
+                mapped = mapping[func_name]
+                if isinstance(mapped, BaseTool):
+                    original_tool = mapped
+            if original_tool is None:
+                for tool in self.original_tools or []:
+                    if sanitize_tool_name(tool.name) == func_name:
+                        original_tool = tool
+                        break
+
+            if not original_tool:
+                continue
+
+            if getattr(original_tool, "result_as_answer", False):
+                return False
+            if getattr(original_tool, "max_usage_count", None) is not None:
+                return False
+
+        return True
+
+    def _execute_single_native_tool_call(self, tool_call: Any) -> dict[str, Any]:
+        """Execute a single native tool call and return metadata/result."""
+        info = extract_tool_call_info(tool_call)
+        if not info:
+            call_id = (
+                getattr(tool_call, "id", None)
+                or (tool_call.get("id") if isinstance(tool_call, dict) else None)
+                or "unknown"
+            )
+            return {
+                "call_id": call_id,
+                "func_name": "unknown",
+                "result": "Error: Invalid native tool call format",
+                "from_cache": False,
+                "original_tool": None,
+            }
+
+        call_id, func_name, func_args = info
+
+        # Parse arguments
+        parsed_args, parse_error = parse_tool_call_args(func_args, func_name, call_id)
+        if parse_error is not None:
+            return parse_error
+        args_dict: dict[str, Any] = parsed_args or {}
+
+        # Get agent_key for event tracking
+        agent_key = getattr(self.agent, "key", "unknown") if self.agent else "unknown"
+
+        original_tool: BaseTool | None = None
+        mapping = getattr(self, "_tool_name_mapping", None)
+        if mapping and func_name in mapping:
+            mapped = mapping[func_name]
+            if isinstance(mapped, BaseTool):
+                original_tool = mapped
+        if original_tool is None:
+            for tool in self.original_tools or []:
+                if sanitize_tool_name(tool.name) == func_name:
+                    original_tool = tool
+                    break
+
+        # Check if tool has reached max usage count
+        max_usage_reached = False
+        if (
+            original_tool
+            and original_tool.max_usage_count is not None
+            and original_tool.current_usage_count >= original_tool.max_usage_count
+        ):
+            max_usage_reached = True
+
+        # Check cache before executing
+        from_cache = False
+        input_str = json.dumps(args_dict) if args_dict else ""
+        if self.tools_handler and self.tools_handler.cache:
+            cached_result = self.tools_handler.cache.read(
+                tool=func_name, input=input_str
+            )
+            if cached_result is not None:
+                result = (
+                    str(cached_result)
+                    if not isinstance(cached_result, str)
+                    else cached_result
+                )
+                from_cache = True
+
+        # Emit tool usage started event
+        started_at = datetime.now()
+        crewai_event_bus.emit(
+            self,
+            event=ToolUsageStartedEvent(
+                tool_name=func_name,
+                tool_args=args_dict,
+                from_agent=self.agent,
+                from_task=self.task,
+                agent_key=agent_key,
+            ),
+        )
+        error_event_emitted = False
+
+        track_delegation_if_needed(func_name, args_dict, self.task)
+
+        structured_tool: CrewStructuredTool | None = None
+        if original_tool is not None:
+            for structured in self.tools or []:
+                if getattr(structured, "_original_tool", None) is original_tool:
+                    structured_tool = structured
+                    break
+        if structured_tool is None:
+            for structured in self.tools or []:
+                if sanitize_tool_name(structured.name) == func_name:
+                    structured_tool = structured
+                    break
+
+        hook_blocked = False
+        before_hook_context = ToolCallHookContext(
+            tool_name=func_name,
+            tool_input=args_dict,
+            tool=structured_tool,  # type: ignore[arg-type]
+            agent=self.agent,
+            task=self.task,
+            crew=self.crew,
+        )
+        before_hooks = get_before_tool_call_hooks()
+        try:
+            for hook in before_hooks:
+                hook_result = hook(before_hook_context)
+                if hook_result is False:
+                    hook_blocked = True
+                    break
+        except Exception as hook_error:
+            if self.agent.verbose:
+                self._printer.print(
+                    content=f"Error in before_tool_call hook: {hook_error}",
+                    color="red",
+                )
+
+        if hook_blocked:
+            result = f"Tool execution blocked by hook. Tool: {func_name}"
+        elif not from_cache and not max_usage_reached:
+            result = "Tool not found"
+            if func_name in self._available_functions:
+                try:
+                    tool_func = self._available_functions[func_name]
+                    raw_result = tool_func(**args_dict)
+
+                    # Add to cache after successful execution (before string conversion)
+                    if self.tools_handler and self.tools_handler.cache:
+                        should_cache = True
+                        if original_tool:
+                            should_cache = original_tool.cache_function(
+                                args_dict, raw_result
+                            )
+                        if should_cache:
+                            self.tools_handler.cache.add(
+                                tool=func_name, input=input_str, output=raw_result
+                            )
+
+                    # Convert to string for message
+                    result = (
+                        str(raw_result)
+                        if not isinstance(raw_result, str)
+                        else raw_result
+                    )
+                except Exception as e:
+                    result = f"Error executing tool: {e}"
+                    if self.task:
+                        self.task.increment_tools_errors()
+                    # Emit tool usage error event
+                    crewai_event_bus.emit(
+                        self,
+                        event=ToolUsageErrorEvent(
+                            tool_name=func_name,
+                            tool_args=args_dict,
+                            from_agent=self.agent,
+                            from_task=self.task,
+                            agent_key=agent_key,
+                            error=e,
+                        ),
+                    )
+                    error_event_emitted = True
+        elif max_usage_reached and original_tool:
+            # Return error message when max usage limit is reached
+            result = f"Tool '{func_name}' has reached its usage limit of {original_tool.max_usage_count} times and cannot be used anymore."
+
+        # Execute after_tool_call hooks (even if blocked, to allow logging/monitoring)
+        after_hook_context = ToolCallHookContext(
+            tool_name=func_name,
+            tool_input=args_dict,
+            tool=structured_tool,  # type: ignore[arg-type]
+            agent=self.agent,
+            task=self.task,
+            crew=self.crew,
+            tool_result=result,
+        )
+        after_hooks = get_after_tool_call_hooks()
+        try:
+            for after_hook in after_hooks:
+                after_hook_result = after_hook(after_hook_context)
+                if after_hook_result is not None:
+                    result = after_hook_result
+                    after_hook_context.tool_result = result
+        except Exception as hook_error:
+            if self.agent.verbose:
+                self._printer.print(
+                    content=f"Error in after_tool_call hook: {hook_error}",
+                    color="red",
+                )
+
+        if not error_event_emitted:
+            crewai_event_bus.emit(
+                self,
+                event=ToolUsageFinishedEvent(
+                    output=result,
+                    tool_name=func_name,
+                    tool_args=args_dict,
+                    from_agent=self.agent,
+                    from_task=self.task,
+                    agent_key=agent_key,
+                    started_at=started_at,
+                    finished_at=datetime.now(),
+                ),
+            )
+
+        return {
+            "call_id": call_id,
+            "func_name": func_name,
+            "result": result,
+            "from_cache": from_cache,
+            "original_tool": original_tool,
+        }
+
     def _extract_tool_name(self, tool_call: Any) -> str:
         """Extract tool name from various tool call formats."""
         if hasattr(tool_call, "function"):
@@ -954,11 +1178,11 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
     def check_max_iterations(
         self,
     ) -> Literal[
-        "force_final_answer", "continue_reasoning", "continue_reasoning_native"
+        "max_iterations_exceeded", "continue_reasoning", "continue_reasoning_native"
     ]:
         """Check if max iterations reached before proceeding with reasoning."""
         if has_reached_max_iterations(self.state.iterations, self.max_iter):
-            return "force_final_answer"
+            return "max_iterations_exceeded"
         if self.state.use_native_tools:
             return "continue_reasoning_native"
         return "continue_reasoning"
@@ -1075,6 +1299,10 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
             self.state.is_finished = False
             self.state.use_native_tools = False
             self.state.pending_tool_calls = []
+            self.state.plan = None
+            self.state.plan_ready = False
+
+            self._kickoff_input = inputs.get("input", "")
 
             if "system" in self.prompt:
                 prompt = cast("SystemPromptResult", self.prompt)
@@ -1157,6 +1385,10 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
             self.state.is_finished = False
             self.state.use_native_tools = False
             self.state.pending_tool_calls = []
+            self.state.plan = None
+            self.state.plan_ready = False
+
+            self._kickoff_input = inputs.get("input", "")
 
             if "system" in self.prompt:
                 prompt = cast("SystemPromptResult", self.prompt)
@@ -1252,7 +1484,9 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
             formatted_answer: Current agent response.
         """
         if self.step_callback:
-            self.step_callback(formatted_answer)
+            cb_result = self.step_callback(formatted_answer)
+            if inspect.iscoroutine(cb_result):
+                asyncio.run(cb_result)
 
     def _append_message_to_state(
         self, text: str, role: Literal["user", "assistant", "system"] = "assistant"

lib/crewai/src/crewai/flow/flow.py

@@ -10,13 +10,15 @@ import asyncio
 from collections.abc import (
     Callable,
     ItemsView,
+    Iterable,
     Iterator,
     KeysView,
     Sequence,
     ValuesView,
 )
-from concurrent.futures import Future
+from concurrent.futures import Future, ThreadPoolExecutor
 import copy
+import enum
 import inspect
 import logging
 import threading
@@ -27,8 +29,10 @@ from typing import (
     Generic,
     Literal,
     ParamSpec,
+    SupportsIndex,
     TypeVar,
     cast,
+    overload,
 )
 from uuid import uuid4
 
@@ -77,7 +81,12 @@ from crewai.flow.flow_wrappers import (
     StartMethod,
 )
 from crewai.flow.persistence.base import FlowPersistence
-from crewai.flow.types import FlowExecutionData, FlowMethodName, InputHistoryEntry, PendingListenerKey
+from crewai.flow.types import (
+    FlowExecutionData,
+    FlowMethodName,
+    InputHistoryEntry,
+    PendingListenerKey,
+)
 from crewai.flow.utils import (
     _extract_all_methods,
     _extract_all_methods_recursive,
@@ -426,8 +435,7 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
     """
 
     def __init__(self, lst: list[T], lock: threading.Lock) -> None:
-        # Do NOT call super().__init__() -- we don't want to copy data into
-        # the builtin list storage. All access goes through self._list.
+        super().__init__()  # empty builtin list; all access goes through self._list
         self._list = lst
         self._lock = lock
 
@@ -435,11 +443,11 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             self._list.append(item)
 
-    def extend(self, items: list[T]) -> None:
+    def extend(self, items: Iterable[T]) -> None:
         with self._lock:
             self._list.extend(items)
 
-    def insert(self, index: int, item: T) -> None:
+    def insert(self, index: SupportsIndex, item: T) -> None:
         with self._lock:
             self._list.insert(index, item)
 
@@ -447,7 +455,7 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             self._list.remove(item)
 
-    def pop(self, index: int = -1) -> T:
+    def pop(self, index: SupportsIndex = -1) -> T:
         with self._lock:
             return self._list.pop(index)
 
@@ -455,15 +463,23 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             self._list.clear()
 
-    def __setitem__(self, index: int, value: T) -> None:
+    @overload
+    def __setitem__(self, index: SupportsIndex, value: T) -> None: ...
+    @overload
+    def __setitem__(self, index: slice, value: Iterable[T]) -> None: ...
+    def __setitem__(self, index: Any, value: Any) -> None:
         with self._lock:
             self._list[index] = value
 
-    def __delitem__(self, index: int) -> None:
+    def __delitem__(self, index: SupportsIndex | slice) -> None:
         with self._lock:
             del self._list[index]
 
-    def __getitem__(self, index: int) -> T:
+    @overload
+    def __getitem__(self, index: SupportsIndex) -> T: ...
+    @overload
+    def __getitem__(self, index: slice) -> list[T]: ...
+    def __getitem__(self, index: Any) -> Any:
         return self._list[index]
 
     def __len__(self) -> int:
@@ -481,7 +497,7 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
     def __bool__(self) -> bool:
         return bool(self._list)
 
-    def __eq__(self, other: object) -> bool:  # type: ignore[override]
+    def __eq__(self, other: object) -> bool:
         """Compare based on the underlying list contents."""
         if isinstance(other, LockedListProxy):
             # Avoid deadlocks by acquiring locks in a consistent order.
@@ -492,7 +508,7 @@ class LockedListProxy(list, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             return self._list == other
 
-    def __ne__(self, other: object) -> bool:  # type: ignore[override]
+    def __ne__(self, other: object) -> bool:
         return not self.__eq__(other)
 
 
@@ -505,8 +521,7 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
     """
 
     def __init__(self, d: dict[str, T], lock: threading.Lock) -> None:
-        # Do NOT call super().__init__() -- we don't want to copy data into
-        # the builtin dict storage. All access goes through self._dict.
+        super().__init__()  # empty builtin dict; all access goes through self._dict
         self._dict = d
         self._lock = lock
 
@@ -518,11 +533,11 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             del self._dict[key]
 
-    def pop(self, key: str, *default: T) -> T:
+    def pop(self, key: str, *default: T) -> T:  # type: ignore[override]
         with self._lock:
             return self._dict.pop(key, *default)
 
-    def update(self, other: dict[str, T]) -> None:
+    def update(self, other: dict[str, T]) -> None:  # type: ignore[override]
         with self._lock:
             self._dict.update(other)
 
@@ -530,7 +545,7 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             self._dict.clear()
 
-    def setdefault(self, key: str, default: T) -> T:
+    def setdefault(self, key: str, default: T) -> T:  # type: ignore[override]
         with self._lock:
             return self._dict.setdefault(key, default)
 
@@ -546,16 +561,16 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
     def __contains__(self, key: object) -> bool:
         return key in self._dict
 
-    def keys(self) -> KeysView[str]:
+    def keys(self) -> KeysView[str]:  # type: ignore[override]
         return self._dict.keys()
 
-    def values(self) -> ValuesView[T]:
+    def values(self) -> ValuesView[T]:  # type: ignore[override]
         return self._dict.values()
 
-    def items(self) -> ItemsView[str, T]:
+    def items(self) -> ItemsView[str, T]:  # type: ignore[override]
         return self._dict.items()
 
-    def get(self, key: str, default: T | None = None) -> T | None:
+    def get(self, key: str, default: T | None = None) -> T | None:  # type: ignore[override]
         return self._dict.get(key, default)
 
     def __repr__(self) -> str:
@@ -564,7 +579,7 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
     def __bool__(self) -> bool:
         return bool(self._dict)
 
-    def __eq__(self, other: object) -> bool:  # type: ignore[override]
+    def __eq__(self, other: object) -> bool:
         """Compare based on the underlying dict contents."""
         if isinstance(other, LockedDictProxy):
             # Avoid deadlocks by acquiring locks in a consistent order.
@@ -575,7 +590,7 @@ class LockedDictProxy(dict, Generic[T]):  # type: ignore[type-arg]
         with self._lock:
             return self._dict == other
 
-    def __ne__(self, other: object) -> bool:  # type: ignore[override]
+    def __ne__(self, other: object) -> bool:
         return not self.__eq__(other)
 
 
@@ -677,6 +692,7 @@ class FlowMeta(type):
                     condition_type = getattr(
                         attr_value, "__condition_type__", OR_CONDITION
                     )
+
                     if (
                         hasattr(attr_value, "__trigger_condition__")
                         and attr_value.__trigger_condition__ is not None
@@ -737,7 +753,9 @@ class Flow(Generic[T], metaclass=FlowMeta):
     name: str | None = None
     tracing: bool | None = None
     stream: bool = False
-    memory: Any = None  # Memory | MemoryScope | MemorySlice | None; auto-created if not set
+    memory: Any = (
+        None  # Memory | MemoryScope | MemorySlice | None; auto-created if not set
+    )
     input_provider: Any = None  # InputProvider | None; per-flow override for self.ask()
 
     def __class_getitem__(cls: type[Flow[T]], item: type[T]) -> type[Flow[T]]:
@@ -752,6 +770,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
         persistence: FlowPersistence | None = None,
         tracing: bool | None = None,
         suppress_flow_events: bool = False,
+        max_method_calls: int = 100,
         **kwargs: Any,
     ) -> None:
         """Initialize a new Flow instance.
@@ -760,6 +779,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
             persistence: Optional persistence backend for storing flow states
             tracing: Whether to enable tracing. True=always enable, False=always disable, None=check environment/user settings
             suppress_flow_events: Whether to suppress flow event emissions (internal use)
+            max_method_calls: Maximum times a single method can be called per execution before raising RecursionError
             **kwargs: Additional state values to initialize or override
         """
         # Initialize basic instance attributes
@@ -775,6 +795,8 @@ class Flow(Generic[T], metaclass=FlowMeta):
         self._completed_methods: set[FlowMethodName] = (
             set()
         )  # Track completed methods for reload
+        self._method_call_counts: dict[FlowMethodName, int] = {}
+        self._max_method_calls = max_method_calls
         self._persistence: FlowPersistence | None = persistence
         self._is_execution_resuming: bool = False
         self._event_futures: list[Future[None]] = []
@@ -881,7 +903,8 @@ class Flow(Generic[T], metaclass=FlowMeta):
         """
         if self.memory is None:
             raise ValueError("No memory configured for this flow")
-        return self.memory.extract_memories(content)
+        result: list[str] = self.memory.extract_memories(content)
+        return result
 
     def _mark_or_listener_fired(self, listener_name: FlowMethodName) -> bool:
         """Mark an OR listener as fired atomically.
@@ -1352,8 +1375,10 @@ class Flow(Generic[T], metaclass=FlowMeta):
             ValueError: If structured state model lacks 'id' field
             TypeError: If state is neither BaseModel nor dictionary
         """
+        init_state = self.initial_state
+
         # Handle case where initial_state is None but we have a type parameter
-        if self.initial_state is None and hasattr(self, "_initial_state_t"):
+        if init_state is None and hasattr(self, "_initial_state_t"):
             state_type = self._initial_state_t
             if isinstance(state_type, type):
                 if issubclass(state_type, FlowState):
@@ -1377,12 +1402,12 @@ class Flow(Generic[T], metaclass=FlowMeta):
                     return cast(T, {"id": str(uuid4())})
 
         # Handle case where no initial state is provided
-        if self.initial_state is None:
+        if init_state is None:
             return cast(T, {"id": str(uuid4())})
 
         # Handle case where initial_state is a type (class)
-        if isinstance(self.initial_state, type):
-            state_class: type[T] = self.initial_state
+        if isinstance(init_state, type):
+            state_class = init_state
             if issubclass(state_class, FlowState):
                 return state_class()
             if issubclass(state_class, BaseModel):
@@ -1393,19 +1418,19 @@ class Flow(Generic[T], metaclass=FlowMeta):
                 if not getattr(model_instance, "id", None):
                     object.__setattr__(model_instance, "id", str(uuid4()))
                 return model_instance
-            if self.initial_state is dict:
+            if init_state is dict:
                 return cast(T, {"id": str(uuid4())})
 
         # Handle dictionary instance case
-        if isinstance(self.initial_state, dict):
-            new_state = dict(self.initial_state)  # Copy to avoid mutations
+        if isinstance(init_state, dict):
+            new_state = dict(init_state)  # Copy to avoid mutations
             if "id" not in new_state:
                 new_state["id"] = str(uuid4())
             return cast(T, new_state)
 
         # Handle BaseModel instance case
-        if isinstance(self.initial_state, BaseModel):
-            model = cast(BaseModel, self.initial_state)
+        if isinstance(init_state, BaseModel):
+            model = cast(BaseModel, init_state)
             if not hasattr(model, "id"):
                 raise ValueError("Flow state model must have an 'id' field")
 
@@ -1719,7 +1744,12 @@ class Flow(Generic[T], metaclass=FlowMeta):
         async def _run_flow() -> Any:
             return await self.kickoff_async(inputs, input_files)
 
-        return asyncio.run(_run_flow())
+        try:
+            asyncio.get_running_loop()
+            with ThreadPoolExecutor(max_workers=1) as pool:
+                return pool.submit(asyncio.run, _run_flow()).result()
+        except RuntimeError:
+            return asyncio.run(_run_flow())
 
     async def kickoff_async(
         self,
@@ -1803,6 +1833,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
                 self._method_outputs.clear()
                 self._pending_and_listeners.clear()
                 self._clear_or_listeners()
+                self._method_call_counts.clear()
             else:
                 # Only enter resumption mode if there are completed methods to
                 # replay.  When _completed_methods is empty (e.g. a pure
@@ -2178,6 +2209,8 @@ class Flow(Generic[T], metaclass=FlowMeta):
             from crewai.flow.async_feedback.types import HumanFeedbackPending
 
             if isinstance(e, HumanFeedbackPending):
+                e.context.method_name = method_name
+
                 # Auto-save pending feedback (create default persistence if needed)
                 if self._persistence is None:
                     from crewai.flow.persistence import SQLiteFlowPersistence
@@ -2277,14 +2310,23 @@ class Flow(Generic[T], metaclass=FlowMeta):
                     router_name, router_input, current_triggering_event_id
                 )
                 if router_result:  # Only add non-None results
-                    router_results.append(FlowMethodName(str(router_result)))
+                    router_result_str = (
+                        router_result.value
+                        if isinstance(router_result, enum.Enum)
+                        else str(router_result)
+                    )
+                    router_results.append(FlowMethodName(router_result_str))
                     # If this was a human_feedback router, map the outcome to the feedback
                     if self.last_human_feedback is not None:
-                        router_result_to_feedback[str(router_result)] = (
+                        router_result_to_feedback[router_result_str] = (
                             self.last_human_feedback
                         )
                 current_trigger = (
-                    FlowMethodName(str(router_result))
+                    FlowMethodName(
+                        router_result.value
+                        if isinstance(router_result, enum.Enum)
+                        else str(router_result)
+                    )
                     if router_result is not None
                     else FlowMethodName("")  # Update for next iteration of router chain
                 )
@@ -2533,6 +2575,16 @@ class Flow(Generic[T], metaclass=FlowMeta):
             - Skips execution if method was already completed (e.g., after reload)
             - Catches and logs any exceptions during execution, preventing individual listener failures from breaking the entire flow
         """
+        count = self._method_call_counts.get(listener_name, 0) + 1
+        if count > self._max_method_calls:
+            raise RecursionError(
+                f"Method '{listener_name}' has been called {self._max_method_calls} times in "
+                f"this flow execution, which indicates an infinite loop. "
+                f"This commonly happens when a @listen label matches the "
+                f"method's own name."
+            )
+        self._method_call_counts[listener_name] = count
+
         if listener_name in self._completed_methods:
             if self._is_execution_resuming:
                 # During resumption, skip execution but continue listeners
@@ -2701,7 +2753,10 @@ class Flow(Generic[T], metaclass=FlowMeta):
                     return topic
             ```
         """
-        from concurrent.futures import ThreadPoolExecutor, TimeoutError as FuturesTimeoutError
+        from concurrent.futures import (
+            ThreadPoolExecutor,
+            TimeoutError as FuturesTimeoutError,
+        )
         from datetime import datetime
 
         from crewai.events.types.flow_events import (
@@ -2770,14 +2825,16 @@ class Flow(Generic[T], metaclass=FlowMeta):
             response = None
 
         # Record in history
-        self._input_history.append({
-            "message": message,
-            "response": response,
-            "method_name": method_name,
-            "timestamp": datetime.now(),
-            "metadata": metadata,
-            "response_metadata": response_metadata,
-        })
+        self._input_history.append(
+            {
+                "message": message,
+                "response": response,
+                "method_name": method_name,
+                "timestamp": datetime.now(),
+                "metadata": metadata,
+                "response_metadata": response_metadata,
+            }
+        )
 
         # Emit input received event
         crewai_event_bus.emit(

lib/crewai/src/crewai/lite_agent.py

@@ -2,10 +2,10 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import Callable
-import time
 from functools import wraps
 import inspect
 import json
+import time
 from types import MethodType
 from typing import (
     TYPE_CHECKING,
@@ -49,15 +49,20 @@ from crewai.events.types.agent_events import (
     LiteAgentExecutionErrorEvent,
     LiteAgentExecutionStartedEvent,
 )
+from crewai.events.types.logging_events import AgentLogsExecutionEvent
 from crewai.events.types.memory_events import (
     MemoryRetrievalCompletedEvent,
     MemoryRetrievalFailedEvent,
     MemoryRetrievalStartedEvent,
 )
-from crewai.events.types.logging_events import AgentLogsExecutionEvent
 from crewai.flow.flow_trackable import FlowTrackable
 from crewai.hooks.llm_hooks import get_after_llm_call_hooks, get_before_llm_call_hooks
-from crewai.hooks.types import AfterLLMCallHookType, BeforeLLMCallHookType
+from crewai.hooks.types import (
+    AfterLLMCallHookCallable,
+    AfterLLMCallHookType,
+    BeforeLLMCallHookCallable,
+    BeforeLLMCallHookType,
+)
 from crewai.lite_agent_output import LiteAgentOutput
 from crewai.llm import LLM
 from crewai.llms.base_llm import BaseLLM
@@ -270,11 +275,11 @@ class LiteAgent(FlowTrackable, BaseModel):
     _guardrail: GuardrailCallable | None = PrivateAttr(default=None)
     _guardrail_retry_count: int = PrivateAttr(default=0)
     _callbacks: list[TokenCalcHandler] = PrivateAttr(default_factory=list)
-    _before_llm_call_hooks: list[BeforeLLMCallHookType] = PrivateAttr(
-        default_factory=get_before_llm_call_hooks
+    _before_llm_call_hooks: list[BeforeLLMCallHookType | BeforeLLMCallHookCallable] = (
+        PrivateAttr(default_factory=get_before_llm_call_hooks)
     )
-    _after_llm_call_hooks: list[AfterLLMCallHookType] = PrivateAttr(
-        default_factory=get_after_llm_call_hooks
+    _after_llm_call_hooks: list[AfterLLMCallHookType | AfterLLMCallHookCallable] = (
+        PrivateAttr(default_factory=get_after_llm_call_hooks)
     )
     _memory: Any = PrivateAttr(default=None)
 
@@ -440,12 +445,16 @@ class LiteAgent(FlowTrackable, BaseModel):
         return self.role
 
     @property
-    def before_llm_call_hooks(self) -> list[BeforeLLMCallHookType]:
+    def before_llm_call_hooks(
+        self,
+    ) -> list[BeforeLLMCallHookType | BeforeLLMCallHookCallable]:
         """Get the before_llm_call hooks for this agent."""
         return self._before_llm_call_hooks
 
     @property
-    def after_llm_call_hooks(self) -> list[AfterLLMCallHookType]:
+    def after_llm_call_hooks(
+        self,
+    ) -> list[AfterLLMCallHookType | AfterLLMCallHookCallable]:
         """Get the after_llm_call hooks for this agent."""
         return self._after_llm_call_hooks
 
@@ -482,11 +491,12 @@ class LiteAgent(FlowTrackable, BaseModel):
         # Inject memory tools once if memory is configured (mirrors Agent._prepare_kickoff)
         if self._memory is not None:
             from crewai.tools.memory_tools import create_memory_tools
-            from crewai.utilities.agent_utils import sanitize_tool_name
+            from crewai.utilities.string_utils import sanitize_tool_name
 
             existing_names = {sanitize_tool_name(t.name) for t in self._parsed_tools}
             memory_tools = [
-                mt for mt in create_memory_tools(self._memory)
+                mt
+                for mt in create_memory_tools(self._memory)
                 if sanitize_tool_name(mt.name) not in existing_names
             ]
             if memory_tools:
@@ -565,9 +575,10 @@ class LiteAgent(FlowTrackable, BaseModel):
             if memory_block:
                 formatted = self.i18n.slice("memory").format(memory=memory_block)
                 if self._messages and self._messages[0].get("role") == "system":
-                    self._messages[0]["content"] = (
-                        self._messages[0].get("content", "") + "\n\n" + formatted
-                    )
+                    existing_content = self._messages[0].get("content", "")
+                    if not isinstance(existing_content, str):
+                        existing_content = ""
+                    self._messages[0]["content"] = existing_content + "\n\n" + formatted
             crewai_event_bus.emit(
                 self,
                 event=MemoryRetrievalCompletedEvent(
@@ -588,16 +599,12 @@ class LiteAgent(FlowTrackable, BaseModel):
             )
 
     def _save_to_memory(self, output_text: str) -> None:
-        """Extract discrete memories from the run and remember each. No-op if _memory is None."""
-        if self._memory is None:
+        """Extract discrete memories from the run and remember each. No-op if _memory is None or read-only."""
+        if self._memory is None or getattr(self._memory, "_read_only", False):
             return
         input_str = self._get_last_user_content() or "User request"
         try:
-            raw = (
-                f"Input: {input_str}\n"
-                f"Agent: {self.role}\n"
-                f"Result: {output_text}"
-            )
+            raw = f"Input: {input_str}\nAgent: {self.role}\nResult: {output_text}"
             extracted = self._memory.extract_memories(raw)
             if extracted:
                 self._memory.remember_many(extracted, agent_role=self.role)
@@ -622,13 +629,20 @@ class LiteAgent(FlowTrackable, BaseModel):
         )
 
         # Execute the agent using invoke loop
-        agent_finish = self._invoke_loop()
+        active_response_format = response_format or self.response_format
+        agent_finish = self._invoke_loop(response_model=active_response_format)
         if self._memory is not None:
-            self._save_to_memory(agent_finish.output)
+            output_text = (
+                agent_finish.output.model_dump_json()
+                if isinstance(agent_finish.output, BaseModel)
+                else agent_finish.output
+            )
+            self._save_to_memory(output_text)
         formatted_result: BaseModel | None = None
 
-        active_response_format = response_format or self.response_format
-        if active_response_format:
+        if isinstance(agent_finish.output, BaseModel):
+            formatted_result = agent_finish.output
+        elif active_response_format:
             try:
                 model_schema = generate_model_description(active_response_format)
                 schema = json.dumps(model_schema, indent=2)
@@ -660,8 +674,13 @@ class LiteAgent(FlowTrackable, BaseModel):
             usage_metrics = self._token_process.get_summary()
 
         # Create output
+        raw_output = (
+            agent_finish.output.model_dump_json()
+            if isinstance(agent_finish.output, BaseModel)
+            else agent_finish.output
+        )
         output = LiteAgentOutput(
-            raw=agent_finish.output,
+            raw=raw_output,
             pydantic=formatted_result,
             agent_role=self.role,
             usage_metrics=usage_metrics.model_dump() if usage_metrics else None,
@@ -838,10 +857,15 @@ class LiteAgent(FlowTrackable, BaseModel):
 
         return formatted_messages
 
-    def _invoke_loop(self) -> AgentFinish:
+    def _invoke_loop(
+        self, response_model: type[BaseModel] | None = None
+    ) -> AgentFinish:
         """
         Run the agent's thought process until it reaches a conclusion or max iterations.
 
+        Args:
+            response_model: Optional Pydantic model for native structured output.
+
         Returns:
             AgentFinish: The final result of the agent execution.
         """
@@ -870,12 +894,19 @@ class LiteAgent(FlowTrackable, BaseModel):
                         printer=self._printer,
                         from_agent=self,
                         executor_context=self,
+                        response_model=response_model,
                         verbose=self.verbose,
                     )
 
                 except Exception as e:
                     raise e
 
+                if isinstance(answer, BaseModel):
+                    formatted_answer = AgentFinish(
+                        thought="", output=answer, text=answer.model_dump_json()
+                    )
+                    break
+
                 formatted_answer = process_llm_response(
                     cast(str, answer), self.use_stop_words
                 )
@@ -901,7 +932,7 @@ class LiteAgent(FlowTrackable, BaseModel):
                     )
 
                 self._append_message(formatted_answer.text, role="assistant")
-            except OutputParserError as e:  # noqa: PERF203
+            except OutputParserError as e:
                 if self.verbose:
                     self._printer.print(
                         content="Failed to parse LLM output. Retrying...",

lib/crewai/src/crewai/llm.py

@@ -427,7 +427,7 @@ class LLM(BaseLLM):
                 f"installed.\n\n"
                 f"To fix this, either:\n"
                 f"  1. Install LiteLLM for broad model support: "
-                f"uv add litellm\n"
+                f"uv add 'crewai[litellm]'\n"
                 f"or\n"
                 f"pip install litellm\n\n"
                 f"For more details, see: "

lib/crewai/src/crewai/llms/base_llm.py

@@ -26,6 +26,7 @@ from crewai.events.types.llm_events import (
     LLMCallStartedEvent,
     LLMCallType,
     LLMStreamChunkEvent,
+    LLMThinkingChunkEvent,
 )
 from crewai.events.types.tool_usage_events import (
     ToolUsageErrorEvent,
@@ -368,9 +369,6 @@ class BaseLLM(ABC):
         """Emit LLM call started event."""
         from crewai.utilities.serialization import to_serializable
 
-        if not hasattr(crewai_event_bus, "emit"):
-            raise ValueError("crewai_event_bus does not have an emit method") from None
-
         crewai_event_bus.emit(
             self,
             event=LLMCallStartedEvent(
@@ -416,9 +414,6 @@ class BaseLLM(ABC):
         from_agent: Agent | None = None,
     ) -> None:
         """Emit LLM call failed event."""
-        if not hasattr(crewai_event_bus, "emit"):
-            raise ValueError("crewai_event_bus does not have an emit method") from None
-
         crewai_event_bus.emit(
             self,
             event=LLMCallFailedEvent(
@@ -449,9 +444,6 @@ class BaseLLM(ABC):
             call_type: The type of LLM call (LLM_CALL or TOOL_CALL).
             response_id: Unique ID for a particular LLM response, chunks have same response_id.
         """
-        if not hasattr(crewai_event_bus, "emit"):
-            raise ValueError("crewai_event_bus does not have an emit method") from None
-
         crewai_event_bus.emit(
             self,
             event=LLMStreamChunkEvent(
@@ -465,6 +457,32 @@ class BaseLLM(ABC):
             ),
         )
 
+    def _emit_thinking_chunk_event(
+        self,
+        chunk: str,
+        from_task: Task | None = None,
+        from_agent: Agent | None = None,
+        response_id: str | None = None,
+    ) -> None:
+        """Emit thinking/reasoning chunk event from a thinking model.
+
+        Args:
+            chunk: The thinking text content.
+            from_task: The task that initiated the call.
+            from_agent: The agent that initiated the call.
+            response_id: Unique ID for a particular LLM response.
+        """
+        crewai_event_bus.emit(
+            self,
+            event=LLMThinkingChunkEvent(
+                chunk=chunk,
+                from_task=from_task,
+                from_agent=from_agent,
+                response_id=response_id,
+                call_id=get_current_call_id(),
+            ),
+        )
+
     def _handle_tool_execution(
         self,
         function_name: str,

lib/crewai/src/crewai/llms/providers/bedrock/completion.py

@@ -234,7 +234,7 @@ class BedrockCompletion(BaseLLM):
         aws_access_key_id: str | None = None,
         aws_secret_access_key: str | None = None,
         aws_session_token: str | None = None,
-        region_name: str = "us-east-1",
+        region_name: str | None = None,
         temperature: float | None = None,
         max_tokens: int | None = None,
         top_p: float | None = None,
@@ -287,15 +287,6 @@ class BedrockCompletion(BaseLLM):
             **kwargs,
         )
 
-        # Initialize Bedrock client with proper configuration
-        session = Session(
-            aws_access_key_id=aws_access_key_id or os.getenv("AWS_ACCESS_KEY_ID"),
-            aws_secret_access_key=aws_secret_access_key
-            or os.getenv("AWS_SECRET_ACCESS_KEY"),
-            aws_session_token=aws_session_token or os.getenv("AWS_SESSION_TOKEN"),
-            region_name=region_name,
-        )
-
         # Configure client with timeouts and retries following AWS best practices
         config = Config(
             read_timeout=300,
@@ -306,8 +297,12 @@ class BedrockCompletion(BaseLLM):
             tcp_keepalive=True,
         )
 
-        self.client = session.client("bedrock-runtime", config=config)
-        self.region_name = region_name
+        self.region_name = (
+            region_name
+            or os.getenv("AWS_DEFAULT_REGION")
+            or os.getenv("AWS_REGION_NAME")
+            or "us-east-1"
+        )
 
         self.aws_access_key_id = aws_access_key_id or os.getenv("AWS_ACCESS_KEY_ID")
         self.aws_secret_access_key = aws_secret_access_key or os.getenv(
@@ -315,6 +310,16 @@ class BedrockCompletion(BaseLLM):
         )
         self.aws_session_token = aws_session_token or os.getenv("AWS_SESSION_TOKEN")
 
+        # Initialize Bedrock client with proper configuration
+        session = Session(
+            aws_access_key_id=self.aws_access_key_id,
+            aws_secret_access_key=self.aws_secret_access_key,
+            aws_session_token=self.aws_session_token,
+            region_name=self.region_name,
+        )
+
+        self.client = session.client("bedrock-runtime", config=config)
+
         self._async_exit_stack = AsyncExitStack() if AIOBOTOCORE_AVAILABLE else None
         self._async_client_initialized = False
 

lib/crewai/src/crewai/llms/providers/gemini/completion.py

@@ -61,6 +61,7 @@ class GeminiCompletion(BaseLLM):
         interceptor: BaseInterceptor[Any, Any] | None = None,
         use_vertexai: bool | None = None,
         response_format: type[BaseModel] | None = None,
+        thinking_config: types.ThinkingConfig | None = None,
         **kwargs: Any,
     ):
         """Initialize Google Gemini chat completion client.
@@ -93,6 +94,10 @@ class GeminiCompletion(BaseLLM):
                          api_version="v1" is automatically configured.
             response_format: Pydantic model for structured output. Used as default when
                            response_model is not passed to call()/acall() methods.
+            thinking_config: ThinkingConfig for thinking models (gemini-2.5+, gemini-3+).
+                           Controls thought output via include_thoughts, thinking_budget,
+                           and thinking_level. When None, thinking models automatically
+                           get include_thoughts=True so thought content is surfaced.
             **kwargs: Additional parameters
         """
         if interceptor is not None:
@@ -139,6 +144,14 @@ class GeminiCompletion(BaseLLM):
             version_match and float(version_match.group(1)) >= 2.0
         )
 
+        self.thinking_config = thinking_config
+        if (
+            self.thinking_config is None
+            and version_match
+            and float(version_match.group(1)) >= 2.5
+        ):
+            self.thinking_config = types.ThinkingConfig(include_thoughts=True)
+
     @property
     def stop(self) -> list[str]:
         """Get stop sequences sent to the API."""
@@ -520,6 +533,9 @@ class GeminiCompletion(BaseLLM):
         if self.safety_settings:
             config_params["safety_settings"] = self.safety_settings
 
+        if self.thinking_config is not None:
+            config_params["thinking_config"] = self.thinking_config
+
         return types.GenerateContentConfig(**config_params)
 
     def _convert_tools_for_interference(  # type: ignore[override]
@@ -618,9 +634,17 @@ class GeminiCompletion(BaseLLM):
                 function_response_part = types.Part.from_function_response(
                     name=tool_name, response=response_data
                 )
-                contents.append(
-                    types.Content(role="user", parts=[function_response_part])
-                )
+                if (
+                    contents
+                    and contents[-1].role == "user"
+                    and contents[-1].parts
+                    and contents[-1].parts[-1].function_response is not None
+                ):
+                    contents[-1].parts.append(function_response_part)
+                else:
+                    contents.append(
+                        types.Content(role="user", parts=[function_response_part])
+                    )
             elif role == "assistant" and message.get("tool_calls"):
                 raw_parts: list[Any] | None = message.get("raw_tool_call_parts")
                 if raw_parts and all(isinstance(p, types.Part) for p in raw_parts):
@@ -894,7 +918,7 @@ class GeminiCompletion(BaseLLM):
         content = self._extract_text_from_response(response)
 
         effective_response_model = None if self.tools else response_model
-        if not effective_response_model:
+        if not response_model:
             content = self._apply_stop_words(content)
 
         return self._finalize_completion_response(
@@ -931,15 +955,6 @@ class GeminiCompletion(BaseLLM):
         if chunk.usage_metadata:
             usage_data = self._extract_token_usage(chunk)
 
-        if chunk.text:
-            full_response += chunk.text
-            self._emit_stream_chunk_event(
-                chunk=chunk.text,
-                from_task=from_task,
-                from_agent=from_agent,
-                response_id=response_id,
-            )
-
         if chunk.candidates:
             candidate = chunk.candidates[0]
             if candidate.content and candidate.content.parts:
@@ -976,6 +991,21 @@ class GeminiCompletion(BaseLLM):
                             call_type=LLMCallType.TOOL_CALL,
                             response_id=response_id,
                         )
+                    elif part.thought and part.text:
+                        self._emit_thinking_chunk_event(
+                            chunk=part.text,
+                            from_task=from_task,
+                            from_agent=from_agent,
+                            response_id=response_id,
+                        )
+                    elif part.text:
+                        full_response += part.text
+                        self._emit_stream_chunk_event(
+                            chunk=part.text,
+                            from_task=from_task,
+                            from_agent=from_agent,
+                            response_id=response_id,
+                        )
 
         return full_response, function_calls, usage_data
 
@@ -1329,7 +1359,7 @@ class GeminiCompletion(BaseLLM):
         text_parts = [
             part.text
             for part in candidate.content.parts
-            if hasattr(part, "text") and part.text
+            if part.text and not part.thought
         ]
 
         return "".join(text_parts)

lib/crewai/src/crewai/mcp/__init__.py

@@ -18,6 +18,7 @@ from crewai.mcp.filters import (
     create_dynamic_tool_filter,
     create_static_tool_filter,
 )
+from crewai.mcp.tool_resolver import MCPToolResolver
 from crewai.mcp.transports.base import BaseTransport, TransportType
 
 
@@ -28,6 +29,7 @@ __all__ = [
     "MCPServerHTTP",
     "MCPServerSSE",
     "MCPServerStdio",
+    "MCPToolResolver",
     "StaticToolFilter",
     "ToolFilter",
     "ToolFilterContext",

lib/crewai/src/crewai/mcp/client.py

@@ -6,7 +6,7 @@ from contextlib import AsyncExitStack
 from datetime import datetime
 import logging
 import time
-from typing import Any
+from typing import Any, NamedTuple
 
 from typing_extensions import Self
 
@@ -34,6 +34,13 @@ from crewai.mcp.transports.stdio import StdioTransport
 from crewai.utilities.string_utils import sanitize_tool_name
 
 
+class _MCPToolResult(NamedTuple):
+    """Internal result from an MCP tool call, carrying the ``isError`` flag."""
+
+    content: str
+    is_error: bool
+
+
 # MCP Connection timeout constants (in seconds)
 MCP_CONNECTION_TIMEOUT = 30  # Increased for slow servers
 MCP_TOOL_EXECUTION_TIMEOUT = 30
@@ -420,6 +427,7 @@ class MCPClient:
         return [
             {
                 "name": sanitize_tool_name(tool.name),
+                "original_name": tool.name,
                 "description": getattr(tool, "description", ""),
                 "inputSchema": getattr(tool, "inputSchema", {}),
             }
@@ -461,29 +469,46 @@ class MCPClient:
         )
 
         try:
-            result = await self._retry_operation(
+            tool_result: _MCPToolResult = await self._retry_operation(
                 lambda: self._call_tool_impl(tool_name, cleaned_arguments),
                 timeout=self.execution_timeout,
             )
 
-            completed_at = datetime.now()
-            execution_duration_ms = (completed_at - started_at).total_seconds() * 1000
-            crewai_event_bus.emit(
-                self,
-                MCPToolExecutionCompletedEvent(
-                    server_name=server_name,
-                    server_url=server_url,
-                    transport_type=transport_type,
-                    tool_name=tool_name,
-                    tool_args=cleaned_arguments,
-                    result=result,
-                    started_at=started_at,
-                    completed_at=completed_at,
-                    execution_duration_ms=execution_duration_ms,
-                ),
-            )
+            finished_at = datetime.now()
+            execution_duration_ms = (finished_at - started_at).total_seconds() * 1000
 
-            return result
+            if tool_result.is_error:
+                crewai_event_bus.emit(
+                    self,
+                    MCPToolExecutionFailedEvent(
+                        server_name=server_name,
+                        server_url=server_url,
+                        transport_type=transport_type,
+                        tool_name=tool_name,
+                        tool_args=cleaned_arguments,
+                        error=tool_result.content,
+                        error_type="tool_error",
+                        started_at=started_at,
+                        failed_at=finished_at,
+                    ),
+                )
+            else:
+                crewai_event_bus.emit(
+                    self,
+                    MCPToolExecutionCompletedEvent(
+                        server_name=server_name,
+                        server_url=server_url,
+                        transport_type=transport_type,
+                        tool_name=tool_name,
+                        tool_args=cleaned_arguments,
+                        result=tool_result.content,
+                        started_at=started_at,
+                        completed_at=finished_at,
+                        execution_duration_ms=execution_duration_ms,
+                    ),
+                )
+
+            return tool_result.content
         except Exception as e:
             failed_at = datetime.now()
             error_type = (
@@ -564,23 +589,27 @@ class MCPClient:
 
         return cleaned
 
-    async def _call_tool_impl(self, tool_name: str, arguments: dict[str, Any]) -> Any:
+    async def _call_tool_impl(
+        self, tool_name: str, arguments: dict[str, Any]
+    ) -> _MCPToolResult:
         """Internal implementation of call_tool."""
         result = await asyncio.wait_for(
             self.session.call_tool(tool_name, arguments),
             timeout=self.execution_timeout,
         )
 
+        is_error = getattr(result, "isError", False) or False
+
         # Extract result content
         if hasattr(result, "content") and result.content:
             if isinstance(result.content, list) and len(result.content) > 0:
                 content_item = result.content[0]
                 if hasattr(content_item, "text"):
-                    return str(content_item.text)
-                return str(content_item)
-            return str(result.content)
+                    return _MCPToolResult(str(content_item.text), is_error)
+                return _MCPToolResult(str(content_item), is_error)
+            return _MCPToolResult(str(result.content), is_error)
 
-        return str(result)
+        return _MCPToolResult(str(result), is_error)
 
     async def list_prompts(self) -> list[dict[str, Any]]:
         """List available prompts from MCP server.

lib/crewai/src/crewai/mcp/tool_resolver.py

@@ -0,0 +1,592 @@
+"""MCP tool resolution for CrewAI agents.
+
+This module extracts all MCP-related tool resolution logic from the Agent class
+into a standalone MCPToolResolver. It handles three flavours of MCP reference:
+
+  1. Native configs:   MCPServerStdio / MCPServerHTTP / MCPServerSSE objects.
+  2. HTTPS URLs:       e.g. "https://mcp.example.com/api"
+  3. AMP references:   e.g. "notion" or "notion#search" (legacy "crewai-amp:" prefix also works)
+"""
+
+from __future__ import annotations
+
+import asyncio
+import time
+from typing import TYPE_CHECKING, Any, Final, cast
+from urllib.parse import urlparse
+
+from crewai.mcp.client import MCPClient
+from crewai.mcp.config import (
+    MCPServerConfig,
+    MCPServerHTTP,
+    MCPServerSSE,
+    MCPServerStdio,
+)
+from crewai.mcp.transports.http import HTTPTransport
+from crewai.mcp.transports.sse import SSETransport
+from crewai.mcp.transports.stdio import StdioTransport
+
+
+if TYPE_CHECKING:
+    from crewai.tools.base_tool import BaseTool
+    from crewai.utilities.logger import Logger
+
+MCP_CONNECTION_TIMEOUT: Final[int] = 10
+MCP_TOOL_EXECUTION_TIMEOUT: Final[int] = 30
+MCP_DISCOVERY_TIMEOUT: Final[int] = 15
+MCP_MAX_RETRIES: Final[int] = 3
+
+_mcp_schema_cache: dict[str, Any] = {}
+_cache_ttl: Final[int] = 300  # 5 minutes
+
+
+class MCPToolResolver:
+    """Resolves MCP server references / configs into CrewAI ``BaseTool`` instances.
+
+    Typical lifecycle::
+
+        resolver = MCPToolResolver(agent=my_agent, logger=my_agent._logger)
+        tools = resolver.resolve(my_agent.mcps)
+        # … agent executes tasks using *tools* …
+        resolver.cleanup()
+
+    The resolver owns the MCP client connections it creates and is responsible
+    for tearing them down via :meth:`cleanup`.
+    """
+
+    def __init__(self, agent: Any, logger: Logger) -> None:
+        self._agent = agent
+        self._logger = logger
+        self._clients: list[Any] = []
+
+    @property
+    def clients(self) -> list[Any]:
+        return list(self._clients)
+
+    def resolve(self, mcps: list[str | MCPServerConfig]) -> list[BaseTool]:
+        """Convert MCP server references/configs to CrewAI tools."""
+        all_tools: list[BaseTool] = []
+        amp_refs: list[tuple[str, str | None]] = []
+
+        for mcp_config in mcps:
+            if isinstance(mcp_config, str) and mcp_config.startswith("https://"):
+                all_tools.extend(self._resolve_external(mcp_config))
+            elif isinstance(mcp_config, str):
+                amp_refs.append(self._parse_amp_ref(mcp_config))
+            else:
+                tools, client = self._resolve_native(mcp_config)
+                all_tools.extend(tools)
+                if client:
+                    self._clients.append(client)
+
+        if amp_refs:
+            tools, clients = self._resolve_amp(amp_refs)
+            all_tools.extend(tools)
+            self._clients.extend(clients)
+
+        return all_tools
+
+    def cleanup(self) -> None:
+        """Disconnect all MCP client connections."""
+        if not self._clients:
+            return
+
+        async def _disconnect_all() -> None:
+            for client in self._clients:
+                if client and hasattr(client, "connected") and client.connected:
+                    await client.disconnect()
+
+        try:
+            asyncio.run(_disconnect_all())
+        except Exception as e:
+            self._logger.log("error", f"Error during MCP client cleanup: {e}")
+        finally:
+            self._clients.clear()
+
+    @staticmethod
+    def _parse_amp_ref(mcp_config: str) -> tuple[str, str | None]:
+        """Parse an AMP reference into *(slug, optional tool name)*.
+
+        Accepts both bare slugs (``"notion"``, ``"notion#search"``) and the
+        legacy ``"crewai-amp:notion"`` form.
+        """
+        bare = mcp_config.removeprefix("crewai-amp:")
+        slug, _, specific_tool = bare.partition("#")
+        return slug, specific_tool or None
+
+    def _resolve_amp(
+        self, amp_refs: list[tuple[str, str | None]]
+    ) -> tuple[list[BaseTool], list[Any]]:
+        """Fetch AMP configs in bulk and return their tools and clients.
+
+        Resolves each unique slug only once (single connection per server),
+        then applies per-ref tool filters to select specific tools.
+        """
+        from crewai.events.event_bus import crewai_event_bus
+        from crewai.events.types.mcp_events import MCPConfigFetchFailedEvent
+
+        unique_slugs = list(dict.fromkeys(slug for slug, _ in amp_refs))
+        amp_configs_map = self._fetch_amp_mcp_configs(unique_slugs)
+
+        all_tools: list[BaseTool] = []
+        all_clients: list[Any] = []
+
+        resolved_cache: dict[str, tuple[list[BaseTool], Any | None]] = {}
+
+        for slug in unique_slugs:
+            config_dict = amp_configs_map.get(slug)
+            if not config_dict:
+                crewai_event_bus.emit(
+                    self,
+                    MCPConfigFetchFailedEvent(
+                        slug=slug,
+                        error=f"Config for '{slug}' not found. Make sure it is connected in your account.",
+                        error_type="not_connected",
+                    ),
+                )
+                continue
+
+            mcp_server_config = self._build_mcp_config_from_dict(config_dict)
+
+            try:
+                tools, client = self._resolve_native(mcp_server_config)
+                resolved_cache[slug] = (tools, client)
+                if client:
+                    all_clients.append(client)
+            except Exception as e:
+                crewai_event_bus.emit(
+                    self,
+                    MCPConfigFetchFailedEvent(
+                        slug=slug,
+                        error=str(e),
+                        error_type="connection_failed",
+                    ),
+                )
+
+        for slug, specific_tool in amp_refs:
+            cached = resolved_cache.get(slug)
+            if not cached:
+                continue
+
+            slug_tools, _ = cached
+            if specific_tool:
+                all_tools.extend(
+                    t for t in slug_tools if t.name.endswith(f"_{specific_tool}")
+                )
+            else:
+                all_tools.extend(slug_tools)
+
+        return all_tools, all_clients
+
+    def _fetch_amp_mcp_configs(self, slugs: list[str]) -> dict[str, dict[str, Any]]:
+        """Fetch MCP server configurations via CrewAI+ API.
+
+        Sends a GET request to the CrewAI+ mcps/configs endpoint with
+        comma-separated slugs. CrewAI+ proxies the request to crewai-oauth.
+
+        API-level failures return ``{}``; individual slugs will then
+        surface as ``MCPConfigFetchFailedEvent`` in :meth:`_resolve_amp`.
+        """
+        import httpx
+
+        try:
+            from crewai_tools.tools.crewai_platform_tools.misc import (
+                get_platform_integration_token,
+            )
+
+            from crewai.cli.plus_api import PlusAPI
+
+            plus_api = PlusAPI(api_key=get_platform_integration_token())
+            response = plus_api.get_mcp_configs(slugs)
+
+            if response.status_code == 200:
+                configs: dict[str, dict[str, Any]] = response.json().get("configs", {})
+                return configs
+
+            self._logger.log(
+                "debug",
+                f"Failed to fetch MCP configs: HTTP {response.status_code}",
+            )
+            return {}
+
+        except httpx.HTTPError as e:
+            self._logger.log("debug", f"Failed to fetch MCP configs: {e}")
+            return {}
+        except Exception as e:
+            self._logger.log("debug", f"Cannot fetch AMP MCP configs: {e}")
+            return {}
+
+    def _resolve_external(self, mcp_ref: str) -> list[BaseTool]:
+        """Resolve an HTTPS MCP server URL into tools."""
+        from crewai.tools.mcp_tool_wrapper import MCPToolWrapper
+
+        if "#" in mcp_ref:
+            server_url, specific_tool = mcp_ref.split("#", 1)
+        else:
+            server_url, specific_tool = mcp_ref, None
+
+        server_params = {"url": server_url}
+        server_name = self._extract_server_name(server_url)
+
+        try:
+            tool_schemas = self._get_mcp_tool_schemas(server_params)
+
+            if not tool_schemas:
+                self._logger.log(
+                    "warning", f"No tools discovered from MCP server: {server_url}"
+                )
+                return []
+
+            tools = []
+            for tool_name, schema in tool_schemas.items():
+                if specific_tool and tool_name != specific_tool:
+                    continue
+
+                try:
+                    wrapper = MCPToolWrapper(
+                        mcp_server_params=server_params,
+                        tool_name=tool_name,
+                        tool_schema=schema,
+                        server_name=server_name,
+                    )
+                    tools.append(wrapper)
+                except Exception as e:
+                    self._logger.log(
+                        "warning",
+                        f"Failed to create MCP tool wrapper for {tool_name}: {e}",
+                    )
+                    continue
+
+            if specific_tool and not tools:
+                self._logger.log(
+                    "warning",
+                    f"Specific tool '{specific_tool}' not found on MCP server: {server_url}",
+                )
+
+            return cast(list[BaseTool], tools)
+
+        except Exception as e:
+            self._logger.log(
+                "warning", f"Failed to connect to MCP server {server_url}: {e}"
+            )
+            return []
+
+    def _resolve_native(
+        self, mcp_config: MCPServerConfig
+    ) -> tuple[list[BaseTool], Any | None]:
+        """Resolve an ``MCPServerConfig`` into tools, returning the client for cleanup."""
+        from crewai.tools.base_tool import BaseTool
+        from crewai.tools.mcp_native_tool import MCPNativeTool
+
+        transport: StdioTransport | HTTPTransport | SSETransport
+        if isinstance(mcp_config, MCPServerStdio):
+            transport = StdioTransport(
+                command=mcp_config.command,
+                args=mcp_config.args,
+                env=mcp_config.env,
+            )
+            server_name = f"{mcp_config.command}_{'_'.join(mcp_config.args)}"
+        elif isinstance(mcp_config, MCPServerHTTP):
+            transport = HTTPTransport(
+                url=mcp_config.url,
+                headers=mcp_config.headers,
+                streamable=mcp_config.streamable,
+            )
+            server_name = self._extract_server_name(mcp_config.url)
+        elif isinstance(mcp_config, MCPServerSSE):
+            transport = SSETransport(
+                url=mcp_config.url,
+                headers=mcp_config.headers,
+            )
+            server_name = self._extract_server_name(mcp_config.url)
+        else:
+            raise ValueError(f"Unsupported MCP server config type: {type(mcp_config)}")
+
+        client = MCPClient(
+            transport=transport,
+            cache_tools_list=mcp_config.cache_tools_list,
+        )
+
+        async def _setup_client_and_list_tools() -> list[dict[str, Any]]:
+            try:
+                if not client.connected:
+                    await client.connect()
+
+                tools_list = await client.list_tools()
+
+                try:
+                    await client.disconnect()
+                    await asyncio.sleep(0.1)
+                except Exception as e:
+                    self._logger.log("error", f"Error during disconnect: {e}")
+
+                return tools_list
+            except Exception as e:
+                if client.connected:
+                    await client.disconnect()
+                    await asyncio.sleep(0.1)
+                raise RuntimeError(
+                    f"Error during setup client and list tools: {e}"
+                ) from e
+
+        try:
+            try:
+                asyncio.get_running_loop()
+                import concurrent.futures
+
+                with concurrent.futures.ThreadPoolExecutor() as executor:
+                    future = executor.submit(
+                        asyncio.run, _setup_client_and_list_tools()
+                    )
+                    tools_list = future.result()
+            except RuntimeError:
+                try:
+                    tools_list = asyncio.run(_setup_client_and_list_tools())
+                except RuntimeError as e:
+                    error_msg = str(e).lower()
+                    if "cancel scope" in error_msg or "task" in error_msg:
+                        raise ConnectionError(
+                            "MCP connection failed due to event loop cleanup issues. "
+                            "This may be due to authentication errors or server unavailability."
+                        ) from e
+                except asyncio.CancelledError as e:
+                    raise ConnectionError(
+                        "MCP connection was cancelled. This may indicate an authentication "
+                        "error or server unavailability."
+                    ) from e
+
+            if mcp_config.tool_filter:
+                filtered_tools = []
+                for tool in tools_list:
+                    if callable(mcp_config.tool_filter):
+                        try:
+                            from crewai.mcp.filters import ToolFilterContext
+
+                            context = ToolFilterContext(
+                                agent=self._agent,
+                                server_name=server_name,
+                                run_context=None,
+                            )
+                            if mcp_config.tool_filter(context, tool):  # type: ignore[call-arg, arg-type]
+                                filtered_tools.append(tool)
+                        except (TypeError, AttributeError):
+                            if mcp_config.tool_filter(tool):  # type: ignore[call-arg, arg-type]
+                                filtered_tools.append(tool)
+                    else:
+                        filtered_tools.append(tool)
+                tools_list = filtered_tools
+
+            tools = []
+            for tool_def in tools_list:
+                tool_name = tool_def.get("name", "")
+                original_tool_name = tool_def.get("original_name", tool_name)
+                if not tool_name:
+                    continue
+
+                args_schema = None
+                if tool_def.get("inputSchema"):
+                    args_schema = self._json_schema_to_pydantic(
+                        tool_name, tool_def["inputSchema"]
+                    )
+
+                tool_schema = {
+                    "description": tool_def.get("description", ""),
+                    "args_schema": args_schema,
+                }
+
+                try:
+                    native_tool = MCPNativeTool(
+                        mcp_client=client,
+                        tool_name=tool_name,
+                        tool_schema=tool_schema,
+                        server_name=server_name,
+                        original_tool_name=original_tool_name,
+                    )
+                    tools.append(native_tool)
+                except Exception as e:
+                    self._logger.log("error", f"Failed to create native MCP tool: {e}")
+                    continue
+
+            return cast(list[BaseTool], tools), client
+        except Exception as e:
+            if client.connected:
+                asyncio.run(client.disconnect())
+
+            raise RuntimeError(f"Failed to get native MCP tools: {e}") from e
+
+    @staticmethod
+    def _build_mcp_config_from_dict(
+        config_dict: dict[str, Any],
+    ) -> MCPServerConfig:
+        """Convert a config dict from crewai-oauth into an MCPServerConfig."""
+        config_type = config_dict.get("type", "http")
+
+        if config_type == "sse":
+            return MCPServerSSE(
+                url=config_dict["url"],
+                headers=config_dict.get("headers"),
+                cache_tools_list=config_dict.get("cache_tools_list", False),
+            )
+
+        return MCPServerHTTP(
+            url=config_dict["url"],
+            headers=config_dict.get("headers"),
+            streamable=config_dict.get("streamable", True),
+            cache_tools_list=config_dict.get("cache_tools_list", False),
+        )
+
+    @staticmethod
+    def _extract_server_name(server_url: str) -> str:
+        """Extract clean server name from URL for tool prefixing."""
+        parsed = urlparse(server_url)
+        domain = parsed.netloc.replace(".", "_")
+        path = parsed.path.replace("/", "_").strip("_")
+        return f"{domain}_{path}" if path else domain
+
+    def _get_mcp_tool_schemas(
+        self, server_params: dict[str, Any]
+    ) -> dict[str, dict[str, Any]]:
+        """Get tool schemas from MCP server with caching."""
+        server_url = server_params["url"]
+
+        cache_key = server_url
+        current_time = time.time()
+
+        if cache_key in _mcp_schema_cache:
+            cached_data, cache_time = _mcp_schema_cache[cache_key]
+            if current_time - cache_time < _cache_ttl:
+                self._logger.log(
+                    "debug", f"Using cached MCP tool schemas for {server_url}"
+                )
+                return cached_data  # type: ignore[no-any-return]
+
+        try:
+            schemas = asyncio.run(self._get_mcp_tool_schemas_async(server_params))
+            _mcp_schema_cache[cache_key] = (schemas, current_time)
+            return schemas
+        except Exception as e:
+            self._logger.log(
+                "warning", f"Failed to get MCP tool schemas from {server_url}: {e}"
+            )
+            return {}
+
+    async def _get_mcp_tool_schemas_async(
+        self, server_params: dict[str, Any]
+    ) -> dict[str, dict[str, Any]]:
+        """Async implementation of MCP tool schema retrieval."""
+        server_url = server_params["url"]
+        return await self._retry_mcp_discovery(
+            self._discover_mcp_tools_with_timeout, server_url
+        )
+
+    async def _retry_mcp_discovery(
+        self, operation_func: Any, server_url: str
+    ) -> dict[str, dict[str, Any]]:
+        """Retry MCP discovery with exponential backoff."""
+        last_error = None
+
+        for attempt in range(MCP_MAX_RETRIES):
+            result, error, should_retry = await self._attempt_mcp_discovery(
+                operation_func, server_url
+            )
+
+            if result is not None:
+                return result
+
+            if not should_retry:
+                raise RuntimeError(error)
+
+            last_error = error
+            if attempt < MCP_MAX_RETRIES - 1:
+                wait_time = 2**attempt
+                await asyncio.sleep(wait_time)
+
+        raise RuntimeError(
+            f"Failed to discover MCP tools after {MCP_MAX_RETRIES} attempts: {last_error}"
+        )
+
+    @staticmethod
+    async def _attempt_mcp_discovery(
+        operation_func: Any, server_url: str
+    ) -> tuple[dict[str, dict[str, Any]] | None, str, bool]:
+        """Attempt single MCP discovery; returns *(result, error_message, should_retry)*."""
+        try:
+            result = await operation_func(server_url)
+            return result, "", False
+
+        except ImportError:
+            return (
+                None,
+                "MCP library not available. Please install with: pip install mcp",
+                False,
+            )
+
+        except asyncio.TimeoutError:
+            return (
+                None,
+                f"MCP discovery timed out after {MCP_DISCOVERY_TIMEOUT} seconds",
+                True,
+            )
+
+        except Exception as e:
+            error_str = str(e).lower()
+
+            if "authentication" in error_str or "unauthorized" in error_str:
+                return None, f"Authentication failed for MCP server: {e!s}", False
+            if "connection" in error_str or "network" in error_str:
+                return None, f"Network connection failed: {e!s}", True
+            if "json" in error_str or "parsing" in error_str:
+                return None, f"Server response parsing error: {e!s}", True
+            return None, f"MCP discovery error: {e!s}", False
+
+    async def _discover_mcp_tools_with_timeout(
+        self, server_url: str
+    ) -> dict[str, dict[str, Any]]:
+        """Discover MCP tools with timeout wrapper."""
+        return await asyncio.wait_for(
+            self._discover_mcp_tools(server_url), timeout=MCP_DISCOVERY_TIMEOUT
+        )
+
+    async def _discover_mcp_tools(self, server_url: str) -> dict[str, dict[str, Any]]:
+        """Discover tools from an MCP server (HTTPS / streamable-HTTP path)."""
+        from mcp import ClientSession
+        from mcp.client.streamable_http import streamablehttp_client
+
+        from crewai.utilities.string_utils import sanitize_tool_name
+
+        async with streamablehttp_client(server_url) as (read, write, _):
+            async with ClientSession(read, write) as session:
+                await asyncio.wait_for(
+                    session.initialize(), timeout=MCP_CONNECTION_TIMEOUT
+                )
+
+                tools_result = await asyncio.wait_for(
+                    session.list_tools(),
+                    timeout=MCP_DISCOVERY_TIMEOUT - MCP_CONNECTION_TIMEOUT,
+                )
+
+                schemas = {}
+                for tool in tools_result.tools:
+                    args_schema = None
+                    if hasattr(tool, "inputSchema") and tool.inputSchema:
+                        args_schema = self._json_schema_to_pydantic(
+                            sanitize_tool_name(tool.name), tool.inputSchema
+                        )
+
+                    schemas[sanitize_tool_name(tool.name)] = {
+                        "description": getattr(tool, "description", ""),
+                        "args_schema": args_schema,
+                    }
+                return schemas
+
+    @staticmethod
+    def _json_schema_to_pydantic(tool_name: str, json_schema: dict[str, Any]) -> type:
+        """Convert JSON Schema to a Pydantic model for tool arguments."""
+        from crewai.utilities.pydantic_schema_utils import create_model_from_schema
+
+        model_name = f"{tool_name.replace('-', '_').replace(' ', '_')}Schema"
+        return create_model_from_schema(
+            json_schema,
+            model_name=model_name,
+            enrich_descriptions=True,
+        )

lib/crewai/src/crewai/memory/__init__.py

@@ -1,6 +1,14 @@
-"""Memory module: unified Memory with LLM analysis and pluggable storage."""
+"""Memory module: unified Memory with LLM analysis and pluggable storage.
+
+Heavy dependencies are lazily imported so that
+``import crewai`` does not initialise at runtime — critical for
+Celery pre-fork and similar deployment patterns.
+"""
+
+from __future__ import annotations
+
+from typing import Any
 
-from crewai.memory.encoding_flow import EncodingFlow
 from crewai.memory.memory_scope import MemoryScope, MemorySlice
 from crewai.memory.types import (
     MemoryMatch,
@@ -10,7 +18,25 @@ from crewai.memory.types import (
     embed_text,
     embed_texts,
 )
-from crewai.memory.unified_memory import Memory
+
+
+_LAZY_IMPORTS: dict[str, tuple[str, str]] = {
+    "Memory": ("crewai.memory.unified_memory", "Memory"),
+    "EncodingFlow": ("crewai.memory.encoding_flow", "EncodingFlow"),
+}
+
+
+def __getattr__(name: str) -> Any:
+    """Lazily import Memory / EncodingFlow to avoid pulling in lancedb at import time."""
+    if name in _LAZY_IMPORTS:
+        import importlib
+
+        module_path, attr = _LAZY_IMPORTS[name]
+        mod = importlib.import_module(module_path)
+        val = getattr(mod, attr)
+        globals()[name] = val
+        return val
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 
 
 __all__ = [

lib/crewai/src/crewai/memory/memory_scope.py

@@ -145,7 +145,7 @@ class MemoryScope:
 
 
 class MemorySlice:
-    """View over multiple scopes: recall searches all, remember requires explicit scope unless read_only."""
+    """View over multiple scopes: recall searches all, remember is a no-op when read_only."""
 
     def __init__(
         self,
@@ -160,7 +160,7 @@ class MemorySlice:
             memory: The underlying Memory instance.
             scopes: List of scope paths to include.
             categories: Optional category filter for recall.
-            read_only: If True, remember() raises PermissionError.
+            read_only: If True, remember() is a silent no-op.
         """
         self._memory = memory
         self._scopes = [s.rstrip("/") or "/" for s in scopes]
@@ -176,10 +176,10 @@ class MemorySlice:
         importance: float | None = None,
         source: str | None = None,
         private: bool = False,
-    ) -> MemoryRecord:
-        """Remember into an explicit scope. Required when read_only=False."""
+    ) -> MemoryRecord | None:
+        """Remember into an explicit scope. No-op when read_only=True."""
         if self._read_only:
-            raise PermissionError("This MemorySlice is read-only")
+            return None
         return self._memory.remember(
             content,
             scope=scope,

lib/crewai/src/crewai/memory/recall_flow.py

@@ -2,7 +2,6 @@
 
 Implements adaptive-depth retrieval with:
 - LLM query distillation into targeted sub-queries
-- Keyword-driven category filtering
 - Time-based filtering from temporal hints
 - Parallel multi-query, multi-scope search
 - Confidence-based routing with iterative deepening (budget loop)
@@ -37,7 +36,6 @@ class RecallState(BaseModel):
     query: str = ""
     scope: str | None = None
     categories: list[str] | None = None
-    inferred_categories: list[str] = Field(default_factory=list)
     time_cutoff: datetime | None = None
     source: str | None = None
     include_private: bool = False
@@ -82,11 +80,8 @@ class RecallFlow(Flow[RecallState]):
     # ------------------------------------------------------------------
 
     def _merged_categories(self) -> list[str] | None:
-        """Merge caller-supplied and LLM-inferred categories."""
-        merged = list(
-            set((self.state.categories or []) + self.state.inferred_categories)
-        )
-        return merged or None
+        """Return caller-supplied categories, or None if empty."""
+        return self.state.categories or None
 
     def _do_search(self) -> list[dict[str, Any]]:
         """Run parallel search across (embeddings x scopes) with filters.
@@ -212,10 +207,6 @@ class RecallFlow(Flow[RecallState]):
             )
             self.state.query_analysis = analysis
 
-            # Wire keywords -> category filter
-            if analysis.keywords:
-                self.state.inferred_categories = analysis.keywords
-
             # Parse time_filter into a datetime cutoff
             if analysis.time_filter:
                 try:

lib/crewai/src/crewai/memory/storage/lancedb_storage.py

@@ -53,6 +53,7 @@ class LanceDBStorage:
         path: str | Path | None = None,
         table_name: str = "memories",
         vector_dim: int | None = None,
+        compact_every: int = 100,
     ) -> None:
         """Initialize LanceDB storage.
 
@@ -64,6 +65,10 @@ class LanceDBStorage:
             vector_dim: Dimensionality of the embedding vector. When ``None``
                   (default), the dimension is auto-detected from the existing
                   table schema or from the first saved embedding.
+            compact_every: Number of ``save()`` calls between automatic
+                  background compactions.  Each ``save()`` creates one new
+                  fragment file; compaction merges them, keeping query
+                  performance consistent.  Set to 0 to disable.
         """
         if path is None:
             storage_dir = os.environ.get("CREWAI_STORAGE_DIR")
@@ -78,6 +83,22 @@ class LanceDBStorage:
         self._table_name = table_name
         self._db = lancedb.connect(str(self._path))
 
+        # On macOS and Linux the default per-process open-file limit is 256.
+        # A LanceDB table stores one file per fragment (one fragment per save()
+        # call by default).  With hundreds of fragments, a single full-table
+        # scan opens all of them simultaneously, exhausting the limit.
+        # Raise it proactively so scans on large tables never hit OS error 24.
+        try:
+            import resource
+            soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+            if soft < 4096:
+                resource.setrlimit(resource.RLIMIT_NOFILE, (min(hard, 4096), hard))
+        except Exception:  # noqa: S110
+            pass  # Windows or already at the max hard limit — safe to ignore
+
+        self._compact_every = compact_every
+        self._save_count = 0
+
         # Get or create a shared write lock for this database path.
         resolved = str(self._path.resolve())
         with LanceDBStorage._path_locks_guard:
@@ -91,6 +112,11 @@ class LanceDBStorage:
         try:
             self._table: lancedb.table.Table | None = self._db.open_table(self._table_name)
             self._vector_dim: int = self._infer_dim_from_table(self._table)
+            # Best-effort: create the scope index if it doesn't exist yet.
+            self._ensure_scope_index()
+            # Compact in the background if the table has accumulated many
+            # fragments from previous runs (each save() creates one).
+            self._compact_if_needed()
         except Exception:
             self._table = None
             self._vector_dim = vector_dim or 0  # 0 = not yet known
@@ -178,6 +204,56 @@ class LanceDBStorage:
         table.delete("id = '__schema_placeholder__'")
         return table
 
+    def _ensure_scope_index(self) -> None:
+        """Create a BTREE scalar index on the ``scope`` column if not present.
+
+        A scalar index lets LanceDB skip a full table scan when filtering by
+        scope prefix, which is the hot path for ``list_records``,
+        ``get_scope_info``, and ``list_scopes``.  The call is best-effort:
+        if the table is empty or the index already exists the exception is
+        swallowed silently.
+        """
+        if self._table is None:
+            return
+        try:
+            self._table.create_scalar_index("scope", index_type="BTREE", replace=False)
+        except Exception:  # noqa: S110
+            pass  # index already exists, table empty, or unsupported version
+
+    # ------------------------------------------------------------------
+    # Automatic background compaction
+    # ------------------------------------------------------------------
+
+    def _compact_if_needed(self) -> None:
+        """Spawn a background compaction on startup.
+
+        Called whenever an existing table is opened so that fragments
+        accumulated in previous sessions are silently merged before the
+        first query.  ``optimize()`` returns quickly when the table is
+        already compact, so the cost is negligible in the common case.
+        """
+        if self._table is None or self._compact_every <= 0:
+            return
+        self._compact_async()
+
+    def _compact_async(self) -> None:
+        """Fire-and-forget: compact the table in a daemon background thread."""
+        threading.Thread(
+            target=self._compact_safe,
+            daemon=True,
+            name="lancedb-compact",
+        ).start()
+
+    def _compact_safe(self) -> None:
+        """Run ``table.optimize()`` in a background thread, absorbing errors."""
+        try:
+            if self._table is not None:
+                self._table.optimize()
+                # Refresh the scope index so new fragments are covered.
+                self._ensure_scope_index()
+        except Exception:
+            _logger.debug("LanceDB background compaction failed", exc_info=True)
+
     def _ensure_table(self, vector_dim: int | None = None) -> lancedb.table.Table:
         """Return the table, creating it lazily if needed.
 
@@ -239,6 +315,7 @@ class LanceDBStorage:
             if r.embedding and len(r.embedding) > 0:
                 dim = len(r.embedding)
                 break
+        is_new_table = self._table is None
         with self._write_lock:
             self._ensure_table(vector_dim=dim)
             rows = [self._record_to_row(r) for r in records]
@@ -246,6 +323,13 @@ class LanceDBStorage:
                 if r["vector"] is None or len(r["vector"]) != self._vector_dim:
                     r["vector"] = [0.0] * self._vector_dim
             self._retry_write("add", rows)
+        # Create the scope index on the first save so it covers the initial dataset.
+        if is_new_table:
+            self._ensure_scope_index()
+        # Auto-compact every N saves so fragment files don't pile up.
+        self._save_count += 1
+        if self._compact_every > 0 and self._save_count % self._compact_every == 0:
+            self._compact_async()
 
     def update(self, record: MemoryRecord) -> None:
         """Update a record by ID. Preserves created_at, updates last_accessed."""
@@ -261,6 +345,10 @@ class LanceDBStorage:
     def touch_records(self, record_ids: list[str]) -> None:
         """Update last_accessed to now for the given record IDs.
 
+        Uses a single batch ``table.update()`` call instead of N
+        delete-and-re-add cycles, which is both faster and avoids
+        unnecessary write amplification.
+
         Args:
             record_ids: IDs of records to touch.
         """
@@ -268,25 +356,20 @@ class LanceDBStorage:
             return
         with self._write_lock:
             now = datetime.utcnow().isoformat()
-            for rid in record_ids:
-                safe_id = str(rid).replace("'", "''")
-                rows = (
-                    self._table.search([0.0] * self._vector_dim)
-                    .where(f"id = '{safe_id}'")
-                    .limit(1)
-                    .to_list()
-                )
-                if rows:
-                    rows[0]["last_accessed"] = now
-                    self._retry_write("delete", f"id = '{safe_id}'")
-                    self._retry_write("add", [rows[0]])
+            safe_ids = [str(rid).replace("'", "''") for rid in record_ids]
+            ids_expr = ", ".join(f"'{rid}'" for rid in safe_ids)
+            self._retry_write(
+                "update",
+                where=f"id IN ({ids_expr})",
+                values={"last_accessed": now},
+            )
 
     def get_record(self, record_id: str) -> MemoryRecord | None:
         """Return a single record by ID, or None if not found."""
         if self._table is None:
             return None
         safe_id = str(record_id).replace("'", "''")
-        rows = self._table.search([0.0] * self._vector_dim).where(f"id = '{safe_id}'").limit(1).to_list()
+        rows = self._table.search().where(f"id = '{safe_id}'").limit(1).to_list()
         if not rows:
             return None
         return self._row_to_record(rows[0])
@@ -374,13 +457,31 @@ class LanceDBStorage:
             self._retry_write("delete", where_expr)
             return before - self._table.count_rows()
 
-    def _scan_rows(self, scope_prefix: str | None = None, limit: int = _SCAN_ROWS_LIMIT) -> list[dict[str, Any]]:
-        """Scan rows optionally filtered by scope prefix."""
+    def _scan_rows(
+        self,
+        scope_prefix: str | None = None,
+        limit: int = _SCAN_ROWS_LIMIT,
+        columns: list[str] | None = None,
+    ) -> list[dict[str, Any]]:
+        """Scan rows optionally filtered by scope prefix.
+
+        Uses a full table scan (no vector query) so the limit is applied after
+        the scope filter, not to ANN candidates before filtering.
+
+        Args:
+            scope_prefix: Optional scope path prefix to filter by.
+            limit: Maximum number of rows to return (applied after filtering).
+            columns: Optional list of column names to fetch.  Pass only the
+                columns you need for metadata operations to avoid reading the
+                heavy ``vector`` column unnecessarily.
+        """
         if self._table is None:
             return []
-        q = self._table.search([0.0] * self._vector_dim)
+        q = self._table.search()
         if scope_prefix is not None and scope_prefix.strip("/"):
             q = q.where(f"scope LIKE '{scope_prefix.rstrip('/')}%'")
+        if columns is not None:
+            q = q.select(columns)
         return q.limit(limit).to_list()
 
     def list_records(
@@ -406,7 +507,10 @@ class LanceDBStorage:
         prefix = scope if scope != "/" else ""
         if prefix and not prefix.startswith("/"):
             prefix = "/" + prefix
-        rows = self._scan_rows(prefix or None)
+        rows = self._scan_rows(
+            prefix or None,
+            columns=["scope", "categories_str", "created_at"],
+        )
         if not rows:
             return ScopeInfo(
                 path=scope or "/",
@@ -453,7 +557,7 @@ class LanceDBStorage:
     def list_scopes(self, parent: str = "/") -> list[str]:
         parent = parent.rstrip("/") or ""
         prefix = (parent + "/") if parent else "/"
-        rows = self._scan_rows(prefix if prefix != "/" else None)
+        rows = self._scan_rows(prefix if prefix != "/" else None, columns=["scope"])
         children: set[str] = set()
         for row in rows:
             sc = str(row.get("scope", ""))
@@ -465,7 +569,7 @@ class LanceDBStorage:
         return sorted(children)
 
     def list_categories(self, scope_prefix: str | None = None) -> dict[str, int]:
-        rows = self._scan_rows(scope_prefix)
+        rows = self._scan_rows(scope_prefix, columns=["categories_str"])
         counts: dict[str, int] = {}
         for row in rows:
             cat_str = row.get("categories_str") or "[]"
@@ -498,6 +602,21 @@ class LanceDBStorage:
         if prefix:
             self._table.delete(f"scope >= '{prefix}' AND scope < '{prefix}/\uFFFF'")
 
+    def optimize(self) -> None:
+        """Compact the table synchronously and refresh the scope index.
+
+        Under normal usage this is called automatically in the background
+        (every ``compact_every`` saves and on startup when the table is
+        fragmented).  Call this explicitly only when you need the compaction
+        to be complete before the next operation — for example immediately
+        after a large bulk import, before a latency-sensitive recall.
+        It is a no-op if the table does not exist.
+        """
+        if self._table is None:
+            return
+        self._table.optimize()
+        self._ensure_scope_index()
+
     async def asave(self, records: list[MemoryRecord]) -> None:
         self.save(records)
 

lib/crewai/src/crewai/memory/types.py

@@ -87,6 +87,22 @@ class MemoryMatch(BaseModel):
         description="Information the system looked for but could not find.",
     )
 
+    def format(self) -> str:
+        """Format this match as a human-readable string including metadata.
+
+        Returns:
+            A multi-line string with score, content, categories, and non-empty
+            metadata fields.
+        """
+        lines = [f"- (score={self.score:.2f}) {self.record.content}"]
+        if self.record.categories:
+            lines.append(f"  categories: {', '.join(self.record.categories)}")
+        if self.record.metadata:
+            for key, value in self.record.metadata.items():
+                if value is not None:
+                    lines.append(f"  {key}: {value}")
+        return "\n".join(lines)
+
 
 class ScopeInfo(BaseModel):
     """Information about a scope in the memory hierarchy."""
@@ -291,7 +307,7 @@ def embed_text(embedder: Any, text: str) -> list[float]:
         return []
     first = result[0]
     if hasattr(first, "tolist"):
-        return first.tolist()
+        return list(first.tolist())
     if isinstance(first, list):
         return [float(x) for x in first]
     return list(first)

lib/crewai/src/crewai/memory/unified_memory.py

@@ -6,7 +6,7 @@ from concurrent.futures import Future, ThreadPoolExecutor
 from datetime import datetime
 import threading
 import time
-from typing import Any, Literal
+from typing import TYPE_CHECKING, Any, Literal
 
 from crewai.events.event_bus import crewai_event_bus
 from crewai.events.types.memory_events import (
@@ -21,7 +21,6 @@ from crewai.llms.base_llm import BaseLLM
 from crewai.memory.analyze import extract_memories_from_content
 from crewai.memory.recall_flow import RecallFlow
 from crewai.memory.storage.backend import StorageBackend
-from crewai.memory.storage.lancedb_storage import LanceDBStorage
 from crewai.memory.types import (
     MemoryConfig,
     MemoryMatch,
@@ -30,13 +29,20 @@ from crewai.memory.types import (
     compute_composite_score,
     embed_text,
 )
+from crewai.rag.embeddings.factory import build_embedder
+from crewai.rag.embeddings.providers.openai.types import OpenAIProviderSpec
 
 
-def _default_embedder() -> Any:
+if TYPE_CHECKING:
+    from chromadb.utils.embedding_functions.openai_embedding_function import (
+        OpenAIEmbeddingFunction,
+    )
+
+
+def _default_embedder() -> OpenAIEmbeddingFunction:
     """Build default OpenAI embedder for memory."""
-    from crewai.rag.embeddings.factory import build_embedder
-
-    return build_embedder({"provider": "openai", "config": {}})
+    spec: OpenAIProviderSpec = {"provider": "openai", "config": {}}
+    return build_embedder(spec)
 
 
 class Memory:
@@ -88,6 +94,10 @@ class Memory:
         # Queries shorter than this skip LLM analysis (saving ~1-3s).
         # Longer queries (full task descriptions) benefit from LLM distillation.
         query_analysis_threshold: int = 200,
+        # When True, all write operations (remember, remember_many) are silently
+        # skipped. Useful for sharing a read-only view of memory across agents
+        # without any of them persisting new memories.
+        read_only: bool = False,
     ) -> None:
         """Initialize Memory.
 
@@ -107,7 +117,9 @@ class Memory:
             complex_query_threshold: For complex queries, explore deeper below this confidence.
             exploration_budget: Number of LLM-driven exploration rounds during deep recall.
             query_analysis_threshold: Queries shorter than this skip LLM analysis during deep recall.
+            read_only: If True, remember() and remember_many() are silent no-ops.
         """
+        self._read_only = read_only
         self._config = MemoryConfig(
             recency_weight=recency_weight,
             semantic_weight=semantic_weight,
@@ -130,14 +142,15 @@ class Memory:
         self._llm_instance: BaseLLM | None = None if isinstance(llm, str) else llm
         self._embedder_config: Any = embedder
         self._embedder_instance: Any = (
-            embedder if (embedder is not None and not isinstance(embedder, dict)) else None
+            embedder
+            if (embedder is not None and not isinstance(embedder, dict))
+            else None
         )
 
-        # Storage is initialized eagerly (local, no API key needed).
-        if storage == "lancedb":
-            self._storage = LanceDBStorage()
-        elif isinstance(storage, str):
-            self._storage = LanceDBStorage(path=storage)
+        if isinstance(storage, str):
+            from crewai.memory.storage.lancedb_storage import LanceDBStorage
+
+            self._storage = LanceDBStorage() if storage == "lancedb" else LanceDBStorage(path=storage)
         else:
             self._storage = storage
 
@@ -160,12 +173,17 @@ class Memory:
             from crewai.llm import LLM
 
             try:
-                self._llm_instance = LLM(model=self._llm_config)
+                model_name = (
+                    self._llm_config
+                    if isinstance(self._llm_config, str)
+                    else str(self._llm_config)
+                )
+                self._llm_instance = LLM(model=model_name)
             except Exception as e:
                 raise RuntimeError(
                     f"Memory requires an LLM for analysis but initialization failed: {e}\n\n"
                     "To fix this, do one of the following:\n"
-                    '  - Set OPENAI_API_KEY for the default model (gpt-4o-mini)\n'
+                    "  - Set OPENAI_API_KEY for the default model (gpt-4o-mini)\n"
                     '  - Pass a different model: Memory(llm="anthropic/claude-3-haiku-20240307")\n'
                     '  - Pass any LLM instance: Memory(llm=LLM(model="your-model"))\n'
                     "  - To skip LLM analysis, pass all fields explicitly to remember()\n"
@@ -180,8 +198,6 @@ class Memory:
         if self._embedder_instance is None:
             try:
                 if isinstance(self._embedder_config, dict):
-                    from crewai.rag.embeddings.factory import build_embedder
-
                     self._embedder_instance = build_embedder(self._embedder_config)
                 else:
                     self._embedder_instance = _default_embedder()
@@ -317,7 +333,7 @@ class Memory:
         source: str | None = None,
         private: bool = False,
         agent_role: str | None = None,
-    ) -> MemoryRecord:
+    ) -> MemoryRecord | None:
         """Store a single item in memory (synchronous).
 
         Routes through the same serialized save pool as ``remember_many``
@@ -335,11 +351,13 @@ class Memory:
             agent_role: Optional agent role for event metadata.
 
         Returns:
-            The created MemoryRecord.
+            The created MemoryRecord, or None if this memory is read-only.
 
         Raises:
             Exception: On save failure (events emitted).
         """
+        if self._read_only:
+            return None
         _source_type = "unified_memory"
         try:
             crewai_event_bus.emit(
@@ -356,7 +374,13 @@ class Memory:
             # then immediately wait for the result.
             future = self._submit_save(
                 self._encode_batch,
-                [content], scope, categories, metadata, importance, source, private,
+                [content],
+                scope,
+                categories,
+                metadata,
+                importance,
+                source,
+                private,
             )
             records = future.result()
             record = records[0] if records else None
@@ -420,13 +444,19 @@ class Memory:
         Returns:
             Empty list (records are not available until the background save completes).
         """
-        if not contents:
+        if not contents or self._read_only:
             return []
 
         self._submit_save(
             self._background_encode_batch,
-            contents, scope, categories, metadata,
-            importance, source, private, agent_role,
+            contents,
+            scope,
+            categories,
+            metadata,
+            importance,
+            source,
+            private,
+            agent_role,
         )
         return []
 
@@ -566,14 +596,13 @@ class Memory:
                     # Privacy filter
                     if not include_private:
                         raw = [
-                            (r, s) for r, s in raw
+                            (r, s)
+                            for r, s in raw
                             if not r.private or r.source == source
                         ]
                     results = []
                     for r, s in raw:
-                        composite, reasons = compute_composite_score(
-                            r, s, self._config
-                        )
+                        composite, reasons = compute_composite_score(r, s, self._config)
                         results.append(
                             MemoryMatch(
                                 record=r,
@@ -739,7 +768,9 @@ class Memory:
             limit: Maximum number of records to return.
             offset: Number of records to skip (for pagination).
         """
-        return self._storage.list_records(scope_prefix=scope, limit=limit, offset=offset)
+        return self._storage.list_records(
+            scope_prefix=scope, limit=limit, offset=offset
+        )
 
     def info(self, path: str = "/") -> ScopeInfo:
         """Return scope info for path."""
@@ -781,7 +812,7 @@ class Memory:
         importance: float | None = None,
         source: str | None = None,
         private: bool = False,
-    ) -> MemoryRecord:
+    ) -> MemoryRecord | None:
         """Async remember: delegates to sync for now."""
         return self.remember(
             content,

lib/crewai/src/crewai/rag/embeddings/factory.py

@@ -216,6 +216,10 @@ def build_embedder_from_dict(
 def build_embedder_from_dict(spec: ONNXProviderSpec) -> ONNXMiniLM_L6_V2: ...
 
 
+@overload
+def build_embedder_from_dict(spec: dict[str, Any]) -> EmbeddingFunction[Any]: ...
+
+
 def build_embedder_from_dict(spec):  # type: ignore[no-untyped-def]
     """Build an embedding function instance from a dictionary specification.
 
@@ -341,6 +345,10 @@ def build_embedder(spec: Text2VecProviderSpec) -> Text2VecEmbeddingFunction: ...
 def build_embedder(spec: ONNXProviderSpec) -> ONNXMiniLM_L6_V2: ...
 
 
+@overload
+def build_embedder(spec: dict[str, Any]) -> EmbeddingFunction[Any]: ...
+
+
 def build_embedder(spec):  # type: ignore[no-untyped-def]
     """Build an embedding function from either a provider spec or a provider instance.
 

lib/crewai/src/crewai/task.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import asyncio
 from concurrent.futures import Future
 from copy import copy as shallow_copy
 import datetime
@@ -585,16 +586,29 @@ class Task(BaseModel):
 
             self._post_agent_execution(agent)
 
-            if not self._guardrails and not self._guardrail:
+            if isinstance(result, BaseModel):
+                raw = result.model_dump_json()
+                if self.output_pydantic:
+                    pydantic_output = result
+                    json_output = None
+                elif self.output_json:
+                    pydantic_output = None
+                    json_output = result.model_dump()
+                else:
+                    pydantic_output = None
+                    json_output = None
+            elif not self._guardrails and not self._guardrail:
+                raw = result
                 pydantic_output, json_output = self._export_output(result)
             else:
+                raw = result
                 pydantic_output, json_output = None, None
 
             task_output = TaskOutput(
                 name=self.name or self.description,
                 description=self.description,
                 expected_output=self.expected_output,
-                raw=result,
+                raw=raw,
                 pydantic=pydantic_output,
                 json_dict=json_output,
                 agent=agent.role,
@@ -624,11 +638,15 @@ class Task(BaseModel):
             self.end_time = datetime.datetime.now()
 
             if self.callback:
-                self.callback(self.output)
+                cb_result = self.callback(self.output)
+                if inspect.isawaitable(cb_result):
+                    await cb_result
 
             crew = self.agent.crew  # type: ignore[union-attr]
             if crew and crew.task_callback and crew.task_callback != self.callback:
-                crew.task_callback(self.output)
+                cb_result = crew.task_callback(self.output)
+                if inspect.isawaitable(cb_result):
+                    await cb_result
 
             if self.output_file:
                 content = (
@@ -682,16 +700,29 @@ class Task(BaseModel):
 
             self._post_agent_execution(agent)
 
-            if not self._guardrails and not self._guardrail:
+            if isinstance(result, BaseModel):
+                raw = result.model_dump_json()
+                if self.output_pydantic:
+                    pydantic_output = result
+                    json_output = None
+                elif self.output_json:
+                    pydantic_output = None
+                    json_output = result.model_dump()
+                else:
+                    pydantic_output = None
+                    json_output = None
+            elif not self._guardrails and not self._guardrail:
+                raw = result
                 pydantic_output, json_output = self._export_output(result)
             else:
+                raw = result
                 pydantic_output, json_output = None, None
 
             task_output = TaskOutput(
                 name=self.name or self.description,
                 description=self.description,
                 expected_output=self.expected_output,
-                raw=result,
+                raw=raw,
                 pydantic=pydantic_output,
                 json_dict=json_output,
                 agent=agent.role,
@@ -722,11 +753,15 @@ class Task(BaseModel):
             self.end_time = datetime.datetime.now()
 
             if self.callback:
-                self.callback(self.output)
+                cb_result = self.callback(self.output)
+                if inspect.iscoroutine(cb_result):
+                    asyncio.run(cb_result)
 
             crew = self.agent.crew  # type: ignore[union-attr]
             if crew and crew.task_callback and crew.task_callback != self.callback:
-                crew.task_callback(self.output)
+                cb_result = crew.task_callback(self.output)
+                if inspect.iscoroutine(cb_result):
+                    asyncio.run(cb_result)
 
             if self.output_file:
                 content = (

lib/crewai/src/crewai/telemetry/__init__.py

@@ -1,5 +1,4 @@
 from crewai.telemetry.telemetry import Telemetry
 
 
-
 __all__ = ["Telemetry"]