Files
crewAI/lib/crewai/src/crewai/experimental/agent_executor.py
Greyson LaLonde 1e27cf3f0f
Some checks failed
CodeQL Advanced / Analyze (actions) (push) Has been cancelled
CodeQL Advanced / Analyze (python) (push) Has been cancelled
Notify Downstream / notify-downstream (push) Has been cancelled
fix: ensure verbosity flag is applied
2026-01-28 11:52:47 -05:00

1431 lines
52 KiB
Python

from __future__ import annotations
from collections.abc import Callable, Coroutine
from datetime import datetime
import json
import threading
from typing import TYPE_CHECKING, Any, Literal, cast
from uuid import uuid4
from pydantic import BaseModel, Field, GetCoreSchemaHandler
from pydantic_core import CoreSchema, core_schema
from rich.console import Console
from rich.text import Text
from crewai.agents.agent_builder.base_agent_executor_mixin import CrewAgentExecutorMixin
from crewai.agents.parser import (
AgentAction,
AgentFinish,
OutputParserError,
)
from crewai.events.event_bus import crewai_event_bus
from crewai.events.listeners.tracing.utils import (
is_tracing_enabled_in_context,
)
from crewai.events.types.logging_events import (
AgentLogsExecutionEvent,
AgentLogsStartedEvent,
)
from crewai.events.types.tool_usage_events import (
ToolUsageErrorEvent,
ToolUsageFinishedEvent,
ToolUsageStartedEvent,
)
from crewai.flow.flow import Flow, listen, or_, router, start
from crewai.hooks.llm_hooks import (
get_after_llm_call_hooks,
get_before_llm_call_hooks,
)
from crewai.hooks.tool_hooks import (
ToolCallHookContext,
get_after_tool_call_hooks,
get_before_tool_call_hooks,
)
from crewai.hooks.types import AfterLLMCallHookType, BeforeLLMCallHookType
from crewai.utilities.agent_utils import (
convert_tools_to_openai_schema,
enforce_rpm_limit,
extract_tool_call_info,
format_message_for_llm,
get_llm_response,
handle_agent_action_core,
handle_context_length,
handle_max_iterations_exceeded,
handle_output_parser_exception,
handle_unknown_error,
has_reached_max_iterations,
is_context_length_exceeded,
is_inside_event_loop,
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.printer import Printer
from crewai.utilities.string_utils import sanitize_tool_name
from crewai.utilities.tool_utils import execute_tool_and_check_finality
from crewai.utilities.training_handler import CrewTrainingHandler
from crewai.utilities.types import LLMMessage
if TYPE_CHECKING:
from crewai.agent import Agent
from crewai.agents.tools_handler import ToolsHandler
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
class AgentReActState(BaseModel):
"""Structured state for agent ReAct flow execution.
Replaces scattered instance variables with validated immutable state.
Maps to: self.messages, self.iterations, formatted_answer in current executor.
"""
messages: list[LLMMessage] = Field(default_factory=list)
iterations: int = Field(default=0)
current_answer: AgentAction | AgentFinish | None = Field(default=None)
is_finished: bool = Field(default=False)
ask_for_human_input: bool = Field(default=False)
use_native_tools: bool = Field(default=False)
pending_tool_calls: list[Any] = Field(default_factory=list)
class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
"""Agent Executor for both standalone agents and crew-bound agents.
Inherits from:
- Flow[AgentReActState]: Provides flow orchestration capabilities
- CrewAgentExecutorMixin: Provides memory methods (short/long/external term)
This executor can operate in two modes:
- Standalone mode: When crew and task are None (used by Agent.kickoff())
- Crew mode: When crew and task are provided (used by Agent.execute_task())
Note: Multiple instances may be created during agent initialization
(cache setup, RPM controller setup, etc.) but only the final instance
should execute tasks via invoke().
"""
def __init__(
self,
llm: BaseLLM,
agent: Agent,
prompt: SystemPromptResult | StandardPromptResult,
max_iter: int,
tools: list[CrewStructuredTool],
tools_names: str,
stop_words: list[str],
tools_description: str,
tools_handler: ToolsHandler,
task: Task | None = None,
crew: Crew | None = None,
step_callback: Any = None,
original_tools: list[BaseTool] | None = None,
function_calling_llm: BaseLLM | Any | None = None,
respect_context_window: bool = False,
request_within_rpm_limit: Callable[[], bool] | None = None,
callbacks: list[Any] | None = None,
response_model: type[BaseModel] | None = None,
i18n: I18N | None = None,
) -> None:
"""Initialize the flow-based agent executor.
Args:
llm: Language model instance.
agent: Agent to execute.
prompt: Prompt templates.
max_iter: Maximum iterations.
tools: Available tools.
tools_names: Tool names string.
stop_words: Stop word list.
tools_description: Tool descriptions.
tools_handler: Tool handler instance.
task: Optional task to execute (None for standalone agent execution).
crew: Optional crew instance (None for standalone agent execution).
step_callback: Optional step callback.
original_tools: Original tool list.
function_calling_llm: Optional function calling LLM.
respect_context_window: Respect context limits.
request_within_rpm_limit: RPM limit check function.
callbacks: Optional callbacks list.
response_model: Optional Pydantic model for structured outputs.
"""
self._i18n: I18N = i18n or get_i18n()
self.llm = llm
self.task: Task | None = task
self.agent = agent
self.crew: Crew | None = crew
self.prompt = prompt
self.tools = tools
self.tools_names = tools_names
self.stop = stop_words
self.max_iter = max_iter
self.callbacks = callbacks or []
self._printer: Printer = Printer()
self.tools_handler = tools_handler
self.original_tools = original_tools or []
self.step_callback = step_callback
self.tools_description = tools_description
self.function_calling_llm = function_calling_llm
self.respect_context_window = respect_context_window
self.request_within_rpm_limit = request_within_rpm_limit
self.response_model = response_model
self.log_error_after = 3
self._console: Console = Console()
# Error context storage for recovery
self._last_parser_error: OutputParserError | None = None
self._last_context_error: Exception | None = None
# Execution guard to prevent concurrent/duplicate executions
self._execution_lock = threading.Lock()
self._is_executing: bool = False
self._has_been_invoked: bool = False
self._flow_initialized: bool = False
self._instance_id = str(uuid4())[:8]
self.before_llm_call_hooks: list[BeforeLLMCallHookType] = []
self.after_llm_call_hooks: list[AfterLLMCallHookType] = []
self.before_llm_call_hooks.extend(get_before_llm_call_hooks())
self.after_llm_call_hooks.extend(get_after_llm_call_hooks())
if self.llm:
existing_stop = getattr(self.llm, "stop", [])
self.llm.stop = list(
set(
existing_stop + self.stop
if isinstance(existing_stop, list)
else self.stop
)
)
self._state = AgentReActState()
def _ensure_flow_initialized(self) -> None:
"""Ensure Flow.__init__() has been called.
This is deferred from __init__ to prevent FlowCreatedEvent emission
during agent setup when multiple executor instances are created.
Only the instance that actually executes via invoke() will emit events.
"""
if not self._flow_initialized:
current_tracing = is_tracing_enabled_in_context()
# Now call Flow's __init__ which will replace self._state
# with Flow's managed state. Suppress flow events since this is
# an agent executor, not a user-facing flow.
super().__init__(
suppress_flow_events=True,
tracing=current_tracing if current_tracing else None,
)
self._flow_initialized = True
def _check_native_tool_support(self) -> bool:
"""Check if LLM supports native function calling.
Returns:
True if the LLM supports native function calling and tools are available.
"""
return (
hasattr(self.llm, "supports_function_calling")
and callable(getattr(self.llm, "supports_function_calling", None))
and self.llm.supports_function_calling()
and bool(self.original_tools)
)
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 = (
convert_tools_to_openai_schema(self.original_tools)
)
def _is_tool_call_list(self, response: list[Any]) -> bool:
"""Check if a response is a list of tool calls.
Args:
response: The response to check.
Returns:
True if the response appears to be a list of tool calls.
"""
if not response:
return False
first_item = response[0]
# Check for OpenAI-style tool call structure
if hasattr(first_item, "function") or (
isinstance(first_item, dict) and "function" in first_item
):
return True
# Check for Anthropic-style tool call structure (ToolUseBlock)
if (
hasattr(first_item, "type")
and getattr(first_item, "type", None) == "tool_use"
):
return True
if hasattr(first_item, "name") and hasattr(first_item, "input"):
return True
# Check for Bedrock-style tool call structure (dict with name and input keys)
if (
isinstance(first_item, dict)
and "name" in first_item
and "input" in first_item
):
return True
# Check for Gemini-style function call (Part with function_call)
if hasattr(first_item, "function_call") and first_item.function_call:
return True
return False
@property
def use_stop_words(self) -> bool:
"""Check to determine if stop words are being used.
Returns:
bool: True if stop words should be used.
"""
return self.llm.supports_stop_words() if self.llm else False
@property
def state(self) -> AgentReActState:
"""Get state - returns temporary state if Flow not yet initialized.
Flow initialization is deferred to prevent event emission during agent setup.
Returns the temporary state until invoke() is called.
"""
return self._state
@property
def messages(self) -> list[LLMMessage]:
"""Compatibility property for mixin - returns state messages."""
return self._state.messages
@messages.setter
def messages(self, value: list[LLMMessage]) -> None:
"""Set state messages."""
self._state.messages = value
@property
def iterations(self) -> int:
"""Compatibility property for mixin - returns state iterations."""
return self._state.iterations
@iterations.setter
def iterations(self, value: int) -> None:
"""Set state iterations."""
self._state.iterations = value
@start()
def initialize_reasoning(self) -> Literal["initialized"]:
"""Initialize the reasoning flow and emit agent start logs."""
self._show_start_logs()
# Check for native tool support on first iteration
if self.state.iterations == 0:
self.state.use_native_tools = self._check_native_tool_support()
if self.state.use_native_tools:
self._setup_native_tools()
return "initialized"
@listen("force_final_answer")
def force_final_answer(self) -> Literal["agent_finished"]:
"""Force agent to provide final answer when max iterations exceeded."""
formatted_answer = handle_max_iterations_exceeded(
formatted_answer=None,
printer=self._printer,
i18n=self._i18n,
messages=list(self.state.messages),
llm=self.llm,
callbacks=self.callbacks,
verbose=self.agent.verbose,
)
self.state.current_answer = formatted_answer
self.state.is_finished = True
return "agent_finished"
@listen("continue_reasoning")
def call_llm_and_parse(self) -> Literal["parsed", "parser_error", "context_error"]:
"""Execute LLM call with hooks and parse the response.
Returns routing decision based on parsing result.
"""
try:
enforce_rpm_limit(self.request_within_rpm_limit)
answer = get_llm_response(
llm=self.llm,
messages=list(self.state.messages),
callbacks=self.callbacks,
printer=self._printer,
from_task=self.task,
from_agent=self.agent,
response_model=None,
executor_context=self,
verbose=self.agent.verbose,
)
# Parse the LLM response
formatted_answer = process_llm_response(answer, self.use_stop_words)
self.state.current_answer = formatted_answer
if "Final Answer:" in answer and isinstance(formatted_answer, AgentAction):
warning_text = Text()
warning_text.append("⚠️ ", style="yellow bold")
warning_text.append(
f"LLM returned 'Final Answer:' but parsed as AgentAction (tool: {formatted_answer.tool})",
style="yellow",
)
self._console.print(warning_text)
preview_text = Text()
preview_text.append("Answer preview: ", style="yellow")
preview_text.append(f"{answer[:200]}...", style="yellow dim")
self._console.print(preview_text)
return "parsed"
except OutputParserError as e:
# Store error context for recovery
self._last_parser_error = e or OutputParserError(
error="Unknown parser error"
)
return "parser_error"
except Exception as e:
if is_context_length_exceeded(e):
self._last_context_error = e
return "context_error"
if e.__class__.__module__.startswith("litellm"):
raise e
handle_unknown_error(self._printer, e, verbose=self.agent.verbose)
raise
@listen("continue_reasoning_native")
def call_llm_native_tools(
self,
) -> Literal["native_tool_calls", "native_finished", "context_error"]:
"""Execute LLM call with native function calling.
Always calls the LLM so it can read reflection prompts and decide
whether to provide a final answer or request more tools.
Returns routing decision based on whether tool calls or final answer.
"""
try:
# Clear pending tools - LLM will decide what to do next after reading
# the reflection prompt. It can either:
# 1. Return a final answer (string) if it has enough info
# 2. Return tool calls (possibly same ones, or different ones)
self.state.pending_tool_calls.clear()
enforce_rpm_limit(self.request_within_rpm_limit)
# Call LLM with native tools
answer = get_llm_response(
llm=self.llm,
messages=list(self.state.messages),
callbacks=self.callbacks,
printer=self._printer,
tools=self._openai_tools,
available_functions=None,
from_task=self.task,
from_agent=self.agent,
response_model=None,
executor_context=self,
verbose=self.agent.verbose,
)
# Check if the response is a list of tool calls
if isinstance(answer, list) and answer and self._is_tool_call_list(answer):
# Store tool calls for sequential processing
self.state.pending_tool_calls = list(answer)
return "native_tool_calls"
# Text response - this is the final answer
if isinstance(answer, str):
self.state.current_answer = AgentFinish(
thought="",
output=answer,
text=answer,
)
self._invoke_step_callback(self.state.current_answer)
self._append_message_to_state(answer)
return "native_finished"
# Unexpected response type, treat as final answer
self.state.current_answer = AgentFinish(
thought="",
output=str(answer),
text=str(answer),
)
self._invoke_step_callback(self.state.current_answer)
self._append_message_to_state(str(answer))
return "native_finished"
except Exception as e:
if is_context_length_exceeded(e):
self._last_context_error = e
return "context_error"
if e.__class__.__module__.startswith("litellm"):
raise e
handle_unknown_error(self._printer, e, verbose=self.agent.verbose)
raise
@router(call_llm_and_parse)
def route_by_answer_type(self) -> Literal["execute_tool", "agent_finished"]:
"""Route based on whether answer is AgentAction or AgentFinish."""
if isinstance(self.state.current_answer, AgentAction):
return "execute_tool"
return "agent_finished"
@listen("execute_tool")
def execute_tool_action(self) -> Literal["tool_completed", "tool_result_is_final"]:
"""Execute the tool action and handle the result."""
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,
tools=self.tools,
i18n=self._i18n,
agent_key=self.agent.key if self.agent else None,
agent_role=self.agent.role if self.agent else None,
tools_handler=self.tools_handler,
task=self.task,
agent=self.agent,
function_calling_llm=self.function_calling_llm,
crew=self.crew,
)
# Handle agent action and append observation to messages
result = self._handle_agent_action(action, tool_result)
self.state.current_answer = result
# 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",
"content": reasoning_prompt,
}
self.state.messages.append(reasoning_message)
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
@listen("native_tool_calls")
def execute_native_tool(
self,
) -> Literal["native_tool_completed", "tool_result_is_final"]:
"""Execute native tool calls in a batch.
Processes all tools from pending_tool_calls, executes them,
and appends results to the conversation history.
Returns:
"native_tool_completed" normally, or "tool_result_is_final" if
a tool with result_as_answer=True was executed.
"""
if not self.state.pending_tool_calls:
return "native_tool_completed"
# Group all tool calls into a single assistant message
tool_calls_to_report = []
for tool_call in self.state.pending_tool_calls:
info = extract_tool_call_info(tool_call)
if not info:
continue
call_id, func_name, func_args = info
tool_calls_to_report.append(
{
"id": call_id,
"type": "function",
"function": {
"name": func_name,
"arguments": func_args
if isinstance(func_args, str)
else json.dumps(func_args),
},
}
)
if tool_calls_to_report:
assistant_message: LLMMessage = {
"role": "assistant",
"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
)
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
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
)
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,
),
)
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:
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,
),
)
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",
)
# Emit tool usage finished event
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(),
),
)
# Append tool result message
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"Tool {func_name} executed with result{cache_info}: {result[:200]}...",
color="green",
)
if (
original_tool
and hasattr(original_tool, "result_as_answer")
and original_tool.result_as_answer
):
# Set the result as the final answer
self.state.current_answer = AgentFinish(
thought="Tool result is the final answer",
output=result,
text=result,
)
self.state.is_finished = True
return "tool_result_is_final"
# Add reflection prompt once after all tools in the batch
reasoning_prompt = self._i18n.slice("post_tool_reasoning")
reasoning_message: LLMMessage = {
"role": "user",
"content": reasoning_prompt,
}
self.state.messages.append(reasoning_message)
return "native_tool_completed"
def _extract_tool_name(self, tool_call: Any) -> str:
"""Extract tool name from various tool call formats."""
if hasattr(tool_call, "function"):
return sanitize_tool_name(tool_call.function.name)
if hasattr(tool_call, "function_call") and tool_call.function_call:
return sanitize_tool_name(tool_call.function_call.name)
if hasattr(tool_call, "name"):
return sanitize_tool_name(tool_call.name)
if isinstance(tool_call, dict):
func_info = tool_call.get("function", {})
return sanitize_tool_name(
func_info.get("name", "") or tool_call.get("name", "unknown")
)
return "unknown"
@router(execute_native_tool)
def increment_native_and_continue(self) -> Literal["initialized"]:
"""Increment iteration counter after native tool execution."""
self.state.iterations += 1
return "initialized"
@listen("initialized")
def continue_iteration(self) -> Literal["check_iteration"]:
"""Bridge listener that connects iteration loop back to iteration check."""
return "check_iteration"
@router(or_(initialize_reasoning, continue_iteration))
def check_max_iterations(
self,
) -> Literal[
"force_final_answer", "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"
if self.state.use_native_tools:
return "continue_reasoning_native"
return "continue_reasoning"
@router(execute_tool_action)
def increment_and_continue(self) -> Literal["initialized"]:
"""Increment iteration counter and loop back for next iteration."""
self.state.iterations += 1
return "initialized"
@listen(or_("agent_finished", "tool_result_is_final", "native_finished"))
def finalize(self) -> Literal["completed", "skipped"]:
"""Finalize execution and emit completion logs."""
if self.state.current_answer is None:
skip_text = Text()
skip_text.append("⚠️ ", style="yellow bold")
skip_text.append(
"Finalize called but no answer in state - skipping", style="yellow"
)
self._console.print(skip_text)
return "skipped"
if not isinstance(self.state.current_answer, AgentFinish):
skip_text = Text()
skip_text.append("⚠️ ", style="yellow bold")
skip_text.append(
f"Finalize called with {type(self.state.current_answer).__name__} instead of AgentFinish - skipping",
style="yellow",
)
self._console.print(skip_text)
return "skipped"
self.state.is_finished = True
self._show_logs(self.state.current_answer)
return "completed"
@listen("parser_error")
def recover_from_parser_error(self) -> Literal["initialized"]:
"""Recover from output parser errors and retry."""
if not self._last_parser_error:
self.state.iterations += 1
return "initialized"
formatted_answer = handle_output_parser_exception(
e=self._last_parser_error,
messages=list(self.state.messages),
iterations=self.state.iterations,
log_error_after=self.log_error_after,
printer=self._printer,
verbose=self.agent.verbose,
)
if formatted_answer:
self.state.current_answer = formatted_answer
self.state.iterations += 1
return "initialized"
@listen("context_error")
def recover_from_context_length(self) -> Literal["initialized"]:
"""Recover from context length errors and retry."""
handle_context_length(
respect_context_window=self.respect_context_window,
printer=self._printer,
messages=self.state.messages,
llm=self.llm,
callbacks=self.callbacks,
i18n=self._i18n,
verbose=self.agent.verbose,
)
self.state.iterations += 1
return "initialized"
def invoke(
self, inputs: dict[str, Any]
) -> dict[str, Any] | Coroutine[Any, Any, dict[str, Any]]:
"""Execute agent with given inputs.
When called from within an existing event loop (e.g., inside a Flow),
this method returns a coroutine that should be awaited. The Flow
framework handles this automatically.
Args:
inputs: Input dictionary containing prompt variables.
Returns:
Dictionary with agent output, or a coroutine if inside an event loop.
"""
# Magic auto-async: if inside event loop, return coroutine for Flow to await
if is_inside_event_loop():
return self.invoke_async(inputs)
self._ensure_flow_initialized()
with self._execution_lock:
if self._is_executing:
raise RuntimeError(
"Executor is already running. "
"Cannot invoke the same executor instance concurrently."
)
self._is_executing = True
self._has_been_invoked = True
try:
# Reset state for fresh execution
self.state.messages.clear()
self.state.iterations = 0
self.state.current_answer = None
self.state.is_finished = False
self.state.use_native_tools = False
self.state.pending_tool_calls = []
if "system" in self.prompt:
prompt = cast("SystemPromptResult", self.prompt)
system_prompt = self._format_prompt(prompt["system"], inputs)
user_prompt = self._format_prompt(prompt["user"], inputs)
self.state.messages.append(
format_message_for_llm(system_prompt, role="system")
)
self.state.messages.append(format_message_for_llm(user_prompt))
else:
user_prompt = self._format_prompt(self.prompt["prompt"], inputs)
self.state.messages.append(format_message_for_llm(user_prompt))
self._inject_files_from_inputs(inputs)
self.state.ask_for_human_input = bool(
inputs.get("ask_for_human_input", False)
)
self.kickoff()
formatted_answer = self.state.current_answer
if not isinstance(formatted_answer, AgentFinish):
raise RuntimeError(
"Agent execution ended without reaching a final answer."
)
if self.state.ask_for_human_input:
formatted_answer = self._handle_human_feedback(formatted_answer)
self._create_short_term_memory(formatted_answer)
self._create_long_term_memory(formatted_answer)
self._create_external_memory(formatted_answer)
return {"output": formatted_answer.output}
except AssertionError:
fail_text = Text()
fail_text.append("", style="red bold")
fail_text.append(
"Agent failed to reach a final answer. This is likely a bug - please report it.",
style="red",
)
self._console.print(fail_text)
raise
except Exception as e:
handle_unknown_error(self._printer, e, verbose=self.agent.verbose)
raise
finally:
self._is_executing = False
async def invoke_async(self, inputs: dict[str, Any]) -> dict[str, Any]:
"""Execute agent asynchronously with given inputs.
This method is designed for use within async contexts, such as when
the agent is called from within an async Flow method. It uses
kickoff_async() directly instead of running in a separate thread.
Args:
inputs: Input dictionary containing prompt variables.
Returns:
Dictionary with agent output.
"""
self._ensure_flow_initialized()
with self._execution_lock:
if self._is_executing:
raise RuntimeError(
"Executor is already running. "
"Cannot invoke the same executor instance concurrently."
)
self._is_executing = True
self._has_been_invoked = True
try:
# Reset state for fresh execution
self.state.messages.clear()
self.state.iterations = 0
self.state.current_answer = None
self.state.is_finished = False
self.state.use_native_tools = False
self.state.pending_tool_calls = []
if "system" in self.prompt:
prompt = cast("SystemPromptResult", self.prompt)
system_prompt = self._format_prompt(prompt["system"], inputs)
user_prompt = self._format_prompt(prompt["user"], inputs)
self.state.messages.append(
format_message_for_llm(system_prompt, role="system")
)
self.state.messages.append(format_message_for_llm(user_prompt))
else:
user_prompt = self._format_prompt(self.prompt["prompt"], inputs)
self.state.messages.append(format_message_for_llm(user_prompt))
self._inject_files_from_inputs(inputs)
self.state.ask_for_human_input = bool(
inputs.get("ask_for_human_input", False)
)
# Use async kickoff directly since we're already in an async context
await self.kickoff_async()
formatted_answer = self.state.current_answer
if not isinstance(formatted_answer, AgentFinish):
raise RuntimeError(
"Agent execution ended without reaching a final answer."
)
if self.state.ask_for_human_input:
formatted_answer = self._handle_human_feedback(formatted_answer)
self._create_short_term_memory(formatted_answer)
self._create_long_term_memory(formatted_answer)
self._create_external_memory(formatted_answer)
return {"output": formatted_answer.output}
except AssertionError:
fail_text = Text()
fail_text.append("", style="red bold")
fail_text.append(
"Agent failed to reach a final answer. This is likely a bug - please report it.",
style="red",
)
self._console.print(fail_text)
raise
except Exception as e:
handle_unknown_error(self._printer, e, verbose=self.agent.verbose)
raise
finally:
self._is_executing = False
async def ainvoke(self, inputs: dict[str, Any]) -> dict[str, Any]:
"""Async version of invoke. Alias for invoke_async."""
return await self.invoke_async(inputs)
def _handle_agent_action(
self, formatted_answer: AgentAction, tool_result: ToolResult
) -> AgentAction | AgentFinish:
"""Process agent action and tool execution result.
Args:
formatted_answer: Agent's action to execute.
tool_result: Result from tool execution.
Returns:
Updated action or final answer.
"""
add_image_tool = self._i18n.tools("add_image")
if (
isinstance(add_image_tool, dict)
and formatted_answer.tool.casefold().strip()
== add_image_tool.get("name", "").casefold().strip()
):
self.state.messages.append(
{"role": "assistant", "content": tool_result.result}
)
return formatted_answer
return handle_agent_action_core(
formatted_answer=formatted_answer,
tool_result=tool_result,
messages=self.state.messages,
step_callback=self.step_callback,
show_logs=self._show_logs,
)
def _invoke_step_callback(
self, formatted_answer: AgentAction | AgentFinish
) -> None:
"""Invoke step callback if configured.
Args:
formatted_answer: Current agent response.
"""
if self.step_callback:
self.step_callback(formatted_answer)
def _append_message_to_state(
self, text: str, role: Literal["user", "assistant", "system"] = "assistant"
) -> None:
"""Add message to state conversation history.
Args:
text: Message content.
role: Message role (default: assistant).
"""
self.state.messages.append(format_message_for_llm(text, role=role))
def _show_start_logs(self) -> None:
"""Emit agent start event."""
if self.agent is None:
raise ValueError("Agent cannot be None")
if self.task is None:
return
crewai_event_bus.emit(
self.agent,
AgentLogsStartedEvent(
agent_role=self.agent.role,
task_description=self.task.description,
verbose=self.agent.verbose
or (hasattr(self, "crew") and getattr(self.crew, "verbose", False)),
),
)
def _show_logs(self, formatted_answer: AgentAction | AgentFinish) -> None:
"""Emit agent execution event.
Args:
formatted_answer: Agent's response to log.
"""
if self.agent is None:
raise ValueError("Agent cannot be None")
crewai_event_bus.emit(
self.agent,
AgentLogsExecutionEvent(
agent_role=self.agent.role,
formatted_answer=formatted_answer,
verbose=self.agent.verbose
or (hasattr(self, "crew") and getattr(self.crew, "verbose", False)),
),
)
def _handle_crew_training_output(
self, result: AgentFinish, human_feedback: str | None = None
) -> None:
"""Save training data for crew training mode.
Args:
result: Agent's final output.
human_feedback: Optional feedback from human.
"""
# Early return if no crew (standalone mode)
if self.crew is None:
return
agent_id = str(self.agent.id)
train_iteration = getattr(self.crew, "_train_iteration", None)
if train_iteration is None or not isinstance(train_iteration, int):
train_error = Text()
train_error.append("", style="red bold")
train_error.append(
"Invalid or missing train iteration. Cannot save training data.",
style="red",
)
self._console.print(train_error)
return
training_handler = CrewTrainingHandler(TRAINING_DATA_FILE)
training_data = training_handler.load() or {}
# Initialize or retrieve agent's training data
agent_training_data = training_data.get(agent_id, {})
if human_feedback is not None:
# Save initial output and human feedback
agent_training_data[train_iteration] = {
"initial_output": result.output,
"human_feedback": human_feedback,
}
else:
# Save improved output
if train_iteration in agent_training_data:
agent_training_data[train_iteration]["improved_output"] = result.output
else:
train_error = Text()
train_error.append("", style="red bold")
train_error.append(
f"No existing training data for agent {agent_id} and iteration "
f"{train_iteration}. Cannot save improved output.",
style="red",
)
self._console.print(train_error)
return
# Update the training data and save
training_data[agent_id] = agent_training_data
training_handler.save(training_data)
def _inject_files_from_inputs(self, inputs: dict[str, Any]) -> None:
"""Inject files from inputs into the last user message.
Args:
inputs: Input dictionary that may contain a 'files' key.
"""
files = inputs.get("files")
if not files:
return
for i in range(len(self.state.messages) - 1, -1, -1):
msg = self.state.messages[i]
if msg.get("role") == "user":
msg["files"] = files
break
@staticmethod
def _format_prompt(prompt: str, inputs: dict[str, str]) -> str:
"""Format prompt template with input values.
Args:
prompt: Template string.
inputs: Values to substitute.
Returns:
Formatted prompt.
"""
prompt = prompt.replace("{input}", inputs["input"])
prompt = prompt.replace("{tool_names}", inputs["tool_names"])
return prompt.replace("{tools}", inputs["tools"])
def _handle_human_feedback(self, formatted_answer: AgentFinish) -> AgentFinish:
"""Process human feedback and refine answer.
Args:
formatted_answer: Initial agent result.
Returns:
Final answer after feedback.
"""
human_feedback = self._ask_human_input(formatted_answer.output)
if self._is_training_mode():
return self._handle_training_feedback(formatted_answer, human_feedback)
return self._handle_regular_feedback(formatted_answer, human_feedback)
def _is_training_mode(self) -> bool:
"""Check if training mode is active.
Returns:
True if in training mode.
"""
return bool(self.crew and self.crew._train)
def _handle_training_feedback(
self, initial_answer: AgentFinish, feedback: str
) -> AgentFinish:
"""Process training feedback and generate improved answer.
Args:
initial_answer: Initial agent output.
feedback: Training feedback.
Returns:
Improved answer.
"""
self._handle_crew_training_output(initial_answer, feedback)
self.state.messages.append(
format_message_for_llm(
self._i18n.slice("feedback_instructions").format(feedback=feedback)
)
)
# Re-run flow for improved answer
self.state.iterations = 0
self.state.is_finished = False
self.state.current_answer = None
self.kickoff()
# Get improved answer from state
improved_answer = self.state.current_answer
if not isinstance(improved_answer, AgentFinish):
raise RuntimeError(
"Training feedback iteration did not produce final answer"
)
self._handle_crew_training_output(improved_answer)
self.state.ask_for_human_input = False
return improved_answer
def _handle_regular_feedback(
self, current_answer: AgentFinish, initial_feedback: str
) -> AgentFinish:
"""Process regular feedback iteratively until user is satisfied.
Args:
current_answer: Current agent output.
initial_feedback: Initial user feedback.
Returns:
Final answer after iterations.
"""
feedback = initial_feedback
answer = current_answer
while self.state.ask_for_human_input:
if feedback.strip() == "":
self.state.ask_for_human_input = False
else:
answer = self._process_feedback_iteration(feedback)
feedback = self._ask_human_input(answer.output)
return answer
def _process_feedback_iteration(self, feedback: str) -> AgentFinish:
"""Process a single feedback iteration and generate updated response.
Args:
feedback: User feedback.
Returns:
Updated agent response.
"""
self.state.messages.append(
format_message_for_llm(
self._i18n.slice("feedback_instructions").format(feedback=feedback)
)
)
# Re-run flow
self.state.iterations = 0
self.state.is_finished = False
self.state.current_answer = None
self.kickoff()
# Get answer from state
answer = self.state.current_answer
if not isinstance(answer, AgentFinish):
raise RuntimeError("Feedback iteration did not produce final answer")
return answer
@classmethod
def __get_pydantic_core_schema__(
cls, _source_type: Any, _handler: GetCoreSchemaHandler
) -> CoreSchema:
"""Generate Pydantic core schema for Protocol compatibility.
Allows the executor to be used in Pydantic models without
requiring arbitrary_types_allowed=True.
"""
return core_schema.any_schema()
# Backward compatibility alias (deprecated)
CrewAgentExecutorFlow = AgentExecutor