husaier · husaier · Aug 26, 2025
diff --git a/examples/agent_patterns/README.md b/examples/agent_patterns/README.md
@@ -37,6 +37,15 @@ For example, you could use an LLM to generate an outline for a story, and then u
 
 See the [`llm_as_a_judge.py`](./llm_as_a_judge.py) file for an example of this.
 
+## Self-reflection and fact checking
+
+Another way to improve response quality is to pair a researcher agent with a critic
+agent. The researcher uses tools such as web search to answer the user's question.
+The critic inspects the answer for unsupported claims and provides feedback. The
+researcher then revises the answer until the critic is satisfied.
+
+See the [`fact_checking.py`](./fact_checking.py) file for an example of this.
+
 ## Parallelization
 
 Running multiple agents in parallel is a common pattern. This can be useful for both latency (e.g. if you have multiple steps that don't depend on each other) and also for other reasons e.g. generating multiple responses and picking the best one.

diff --git a/examples/agent_patterns/fact_checking.py b/examples/agent_patterns/fact_checking.py
@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass
+from typing import Literal
+
+from agents import Agent, ItemHelpers, Runner, TResponseInputItem, WebSearchTool, trace
+
+"""Example of self-reflection and fact checking using two cooperating agents.
+
+The research agent answers a question using web search. The critic agent reviews the
+answer for unsupported claims. If the critic finds issues, it provides feedback and the
+researcher revises the answer. The loop continues until the critic approves the response.
+"""
+
+researcher = Agent(
+    name="researcher",
+    instructions=(
+        "You answer the user's question by searching the web for facts."
+        "Think step by step and cite your sources."
+        "If feedback is provided, use it to revise your answer."
+    ),
+    tools=[WebSearchTool()],
+)
+
+
+@dataclass
+class CriticFeedback:
+    verdict: Literal["pass", "revise"]
+    feedback: str
+
+
+critic = Agent[None](
+    name="critic",
+    instructions=(
+        "You check the researcher's answer for unsupported or inaccurate claims."
+        "If everything is well supported, respond with verdict 'pass'."
+        "Otherwise, respond with verdict 'revise' and explain what needs fixing."
+    ),
+    output_type=CriticFeedback,
+)
+
+
+async def main() -> None:
+    question = input("Question: ")
+    conversation: list[TResponseInputItem] = [{"role": "user", "content": question}]
+
+    latest_answer = ""
+    with trace("fact_checking_loop"):
+        while True:
+            research_result = await Runner.run(researcher, conversation)
+            conversation = research_result.to_input_list()
+            latest_answer = ItemHelpers.text_message_outputs(research_result.new_items)
+            print("Researcher answer:\n", latest_answer)
+
+            critic_result = await Runner.run(critic, conversation)
+            feedback = critic_result.final_output
+            print(f"Critic verdict: {feedback.verdict}")
+
+            if feedback.verdict == "pass":
+                break
+
+            print("Revising with critic feedback\n")
+            conversation.append({"role": "user", "content": f"Feedback: {feedback.feedback}"})
+
+    print("Final answer:\n", latest_answer)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())