Reflection

Agents make mistakes: wrong tool arguments, misinterpreted results, incomplete answers. Without reflection, these errors propagate to the final output. Reflection adds a self-correction loop — the agent reviews its work before returning to the user.

Reflection is like a writer reviewing their draft before submitting — catching typos, logical gaps, and unclear sections that seemed fine during writing.

Scenario

Agent drafts email to client. Reflection: 'The email mentions Q3 revenue but I never queried the database for Q3 data — I hallucinated the number.' Agent re-enters loop, queries DB, drafts corrected email.

Solution

In Agent Foundations, apply Reflection to this scenario: Agent drafts email to client. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Reflection (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Reflection happens in the code.

1REFLECTION_PROMPT = """Review this agent output for errors:  # key line for Reflection2
3Task: {task}4Agent actions: {actions}5Draft answer: {draft}6
7Check:81. Is every claim supported by tool results?92. Are there hallucinated facts?103. Is the task fully completed?114. What should be improved?12
13Return JSON: {"pass": bool, "issues": [...], "suggestion": "..."}"""14
15def reflect(task, actions, draft) -> dict:  # define a reusable function16    response = client.chat.completions.create(  # call the API17        model="gpt-4o",18        messages=[{"role": "user", "content": REFLECTION_PROMPT.format(  # key line for Reflection19            task=task, actions=actions, draft=draft)}],20        response_format={"type": "json_object"},21        temperature=0,22    )23    return json.loads(response.choices[0].message.content)  # return the result

• Reflection on every task — unnecessary latency and cost
• Self-critique only — misses errors the same model makes
• Not storing reflection insights for future tasks
• Infinite reflection loops — need max retry limit
• Reflecting on format but not factual accuracy

Quick recap — the most important points from this module.