Fix agent loop not stopping after max iterations reached

Fixes #3847

The agent execution loop was not stopping after max_iter was reached.
After calling handle_max_iterations_exceeded(), the loop continued and
made additional LLM calls, overwriting the formatted answer.

Changes:
- Modified handle_max_iterations_exceeded() to always return AgentFinish
  instead of AgentAction | AgentFinish, ensuring a final answer is always
  produced when max iterations are exceeded
- Added early return in _invoke_loop() after handle_max_iterations_exceeded()
  to prevent additional LLM calls
- Updated test_agent_custom_max_iterations to expect 2 LLM calls instead of 3
  (1 initial + 1 from max_iter handler, no longer 3 due to extra call bug)
- Added test_agent_max_iterations_zero to verify no get_llm_response calls
  when max_iter=0
- Added test_agent_max_iterations_one_stops_after_two_calls to verify exactly
  2 LLM calls with max_iter=1
- Fixed type annotations to properly handle AgentAction | AgentFinish | None

Co-Authored-By: João <joao@crewai.com>

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

@@ -202,7 +202,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
         Returns:
             Final answer from the agent.
         """
-        formatted_answer = None
+        formatted_answer: AgentAction | AgentFinish | None = None
         while not isinstance(formatted_answer, AgentFinish):
             try:
                 if has_reached_max_iterations(self.iterations, self.max_iter):
@@ -214,6 +214,8 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                         llm=self.llm,
                         callbacks=self.callbacks,
                     )
+                    self._show_logs(formatted_answer)
+                    return formatted_answer
 
                 enforce_rpm_limit(self.request_within_rpm_limit)
 
@@ -257,9 +259,12 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                     formatted_answer = self._handle_agent_action(
                         formatted_answer, tool_result
                     )
-
-                self._invoke_step_callback(formatted_answer)
-                self._append_message(formatted_answer.text)
+                    
+                    self._invoke_step_callback(formatted_answer)
+                    self._append_message(formatted_answer.text)
+                elif isinstance(formatted_answer, AgentFinish):
+                    self._invoke_step_callback(formatted_answer)
+                    self._append_message(formatted_answer.text)
 
             except OutputParserError as e:  # noqa: PERF203
                 formatted_answer = handle_output_parser_exception(

lib/crewai/src/crewai/utilities/agent_utils.py

@@ -127,7 +127,7 @@ def handle_max_iterations_exceeded(
     messages: list[LLMMessage],
     llm: LLM | BaseLLM,
     callbacks: list[TokenCalcHandler],
-) -> AgentAction | AgentFinish:
+) -> AgentFinish:
     """Handles the case when the maximum number of iterations is exceeded. Performs one more LLM call to get the final answer.
 
     Args:
@@ -139,7 +139,7 @@ def handle_max_iterations_exceeded(
         callbacks: List of callbacks for the LLM call.
 
     Returns:
-        The final formatted answer after exceeding max iterations.
+        The final formatted answer after exceeding max iterations (always AgentFinish).
     """
     printer.print(
         content="Maximum iterations reached. Requesting final answer.",
@@ -168,8 +168,17 @@ def handle_max_iterations_exceeded(
         )
         raise ValueError("Invalid response from LLM call - None or empty.")
 
-    # Return the formatted answer, regardless of its type
-    return format_answer(answer=answer)
+    # Parse the answer and ensure it's always an AgentFinish
+    parsed_answer = format_answer(answer=answer)
+
+    if isinstance(parsed_answer, AgentAction):
+        return AgentFinish(
+            thought=parsed_answer.thought,
+            output=parsed_answer.text,
+            text=parsed_answer.text,
+        )
+
+    return parsed_answer
 
 
 def format_message_for_llm(

lib/crewai/tests/agents/test_agent.py

@@ -508,7 +508,96 @@ def test_agent_custom_max_iterations():
     assert isinstance(result, str)
     assert len(result) > 0
     assert call_count > 0
-    assert call_count == 3
+    assert call_count == 2
+
+
+def test_agent_max_iterations_zero():
+    """Test that with max_iter=0, get_llm_response is never called but handle_max_iterations_exceeded makes one LLM call."""
+    from unittest.mock import MagicMock
+    from crewai.agents import crew_agent_executor
+
+    agent = Agent(
+        role="test role",
+        goal="test goal",
+        backstory="test backstory",
+        max_iter=0,
+        allow_delegation=False,
+    )
+
+    original_call = agent.llm.call
+    llm_call_count = 0
+
+    def counting_llm_call(*args, **kwargs):
+        nonlocal llm_call_count
+        llm_call_count += 1
+        return original_call(*args, **kwargs)
+
+    agent.llm.call = counting_llm_call
+
+    get_llm_response_call_count = 0
+    original_get_llm_response = crew_agent_executor.get_llm_response
+
+    def counting_get_llm_response(*args, **kwargs):
+        nonlocal get_llm_response_call_count
+        get_llm_response_call_count += 1
+        return original_get_llm_response(*args, **kwargs)
+
+    crew_agent_executor.get_llm_response = counting_get_llm_response
+
+    try:
+        task = Task(
+            description="What is 2+2?",
+            expected_output="The answer",
+        )
+        result = agent.execute_task(task=task)
+
+        assert result is not None
+        assert isinstance(result, str)
+        assert len(result) > 0
+        assert get_llm_response_call_count == 0
+        assert llm_call_count == 1
+    finally:
+        crew_agent_executor.get_llm_response = original_get_llm_response
+
+
+def test_agent_max_iterations_one_stops_after_two_calls():
+    """Test that with max_iter=1, exactly 2 LLM calls are made (initial + max_iter handler)."""
+    @tool
+    def dummy_tool() -> str:
+        """A dummy tool that returns a value."""
+        return "dummy result"
+
+    agent = Agent(
+        role="test role",
+        goal="test goal",
+        backstory="test backstory",
+        max_iter=1,
+        allow_delegation=False,
+    )
+
+    original_call = agent.llm.call
+    llm_call_count = 0
+
+    def counting_llm_call(*args, **kwargs):
+        nonlocal llm_call_count
+        llm_call_count += 1
+        return original_call(*args, **kwargs)
+
+    agent.llm.call = counting_llm_call
+
+    task = Task(
+        description="Keep using the dummy_tool repeatedly.",
+        expected_output="The final answer",
+    )
+    result = agent.execute_task(
+        task=task,
+        tools=[dummy_tool],
+    )
+
+    assert result is not None
+    assert isinstance(result, str)
+    assert len(result) > 0
+    assert llm_call_count == 2
 
 
 @pytest.mark.vcr(filter_headers=["authorization"])