CrewAgentExecutor does not reset messages and iterations between task executions

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

@@ -210,6 +210,10 @@ def invoke(self, inputs: dict[str, Any]) -> dict[str, Any]:
         Returns:
             Dictionary with agent output.
         """
+        # Reset state for fresh execution
+        self.messages.clear()
+        self.iterations = 0
+
         self._setup_messages(inputs)
 
         self._inject_multimodal_files(inputs)
@@ -987,6 +991,10 @@ async def ainvoke(self, inputs: dict[str, Any]) -> dict[str, Any]:
         Returns:
             Dictionary with agent output.
         """
+        # Reset state for fresh execution
+        self.messages.clear()
+        self.iterations = 0
+
         self._setup_messages(inputs)
 
         await self._ainject_multimodal_files(inputs)

lib/crewai/tests/agents/test_async_agent_executor.py

@@ -83,9 +83,17 @@ class TestAsyncAgentExecutor:
 
     @pytest.mark.asyncio
     async def test_ainvoke_returns_output(self, executor: CrewAgentExecutor) -> None:
-        """Test that ainvoke returns the expected output."""
+        """Test that ainvoke returns the expected output and resets stale state."""
         expected_output = "Final answer from agent"
 
+        # Pre-populate stale state from a previous task execution to verify
+        # that ainvoke() resets it before starting the new execution.
+        executor.messages = [
+            {"role": "user", "content": "stale prompt from task 1"},
+            {"role": "assistant", "content": "stale response from task 1"},
+        ]
+        executor.iterations = 5
+
         with patch.object(
             executor,
             "_ainvoke_loop",
@@ -108,6 +116,13 @@ async def test_ainvoke_returns_output(self, executor: CrewAgentExecutor) -> None
 
         assert result == {"output": expected_output}
 
+        # Stale state must have been cleared before execution
+        all_content = " ".join(
+            str(m.get("content", "")) for m in executor.messages
+        )
+        assert "stale" not in all_content
+        assert executor.iterations == 0
+
     @pytest.mark.asyncio
     async def test_ainvoke_loop_calls_aget_llm_response(
         self, executor: CrewAgentExecutor
@@ -294,6 +309,85 @@ async def delayed_response(*args: Any, **kwargs: Any) -> str:
         assert all("output" in r for r in results)
         assert max_concurrent > 1, f"Expected concurrent execution, max concurrent was {max_concurrent}"
 
+    def test_sequential_invoke_produces_isolated_messages(
+        self, executor: CrewAgentExecutor
+    ) -> None:
+        """Test that consecutive invoke() calls on the same executor produce
+        isolated message histories, as happens when Crew reuses an agent
+        for sequential tasks.
+        """
+        # --- First task ---
+        def loop_task1() -> AgentFinish:
+            executor.messages.append(
+                {"role": "assistant", "content": "task1 answer"}
+            )
+            executor.iterations = 3
+            return AgentFinish(
+                thought="", output="Task 1 done", text="Task 1 done"
+            )
+
+        with patch.object(executor, "_invoke_loop", side_effect=loop_task1):
+            with patch.object(executor, "_show_start_logs"):
+                with patch.object(executor, "_create_short_term_memory"):
+                    with patch.object(executor, "_create_long_term_memory"):
+                        with patch.object(executor, "_create_external_memory"):
+                            executor.invoke(
+                                {
+                                    "input": "first task",
+                                    "tool_names": "",
+                                    "tools": "",
+                                }
+                            )
+
+        # Sanity: task 1 state is present
+        assert any(
+            "task1" in str(m.get("content", "")) for m in executor.messages
+        )
+        task1_iterations = executor.iterations
+
+        # --- Second task (same executor, simulating Crew reuse) ---
+        messages_at_loop_start: list[dict[str, Any]] = []
+
+        def loop_task2() -> AgentFinish:
+            messages_at_loop_start.extend(
+                [dict(m) for m in executor.messages]
+            )
+            return AgentFinish(
+                thought="", output="Task 2 done", text="Task 2 done"
+            )
+
+        with patch.object(executor, "_invoke_loop", side_effect=loop_task2):
+            with patch.object(executor, "_show_start_logs"):
+                with patch.object(executor, "_create_short_term_memory"):
+                    with patch.object(executor, "_create_long_term_memory"):
+                        with patch.object(executor, "_create_external_memory"):
+                            executor.invoke(
+                                {
+                                    "input": "second task",
+                                    "tool_names": "",
+                                    "tools": "",
+                                }
+                            )
+
+        # No task 1 messages should be visible when task 2's loop starts
+        for msg in messages_at_loop_start:
+            content = str(msg.get("content", ""))
+            assert "task1" not in content, (
+                f"Task 1 message leaked into task 2: {content}"
+            )
+            assert "first task" not in content, (
+                f"Task 1 prompt leaked into task 2: {content}"
+            )
+
+        # Task 2 should see its own prompt
+        assert any(
+            "second task" in str(m.get("content", ""))
+            for m in messages_at_loop_start
+        )
+
+        # Iterations must not carry over
+        assert executor.iterations < task1_iterations
+
 
 class TestAsyncLLMResponseHelper:
     """Tests for aget_llm_response helper function."""