Reflexion

Standard agents repeat the same mistakes every time. Reflexion adds a learning loop: try → fail → reflect on why → store lesson → retry with lesson in context. Agents get better at specific tasks through experience, not weight updates.

Reflexion is like a student who keeps a mistake journal — after each failed test, they write what went wrong and review the journal before the next attempt.

Scenario

Trial 1: Agent queries 'sales data' without date filter → wrong results. Reflection: 'Always specify date range for sales queries.' Trial 2: Agent queries 'Q3 2024 sales' → correct results. Lesson stored for future sales tasks.

Solution

In Agent Design Patterns, apply Reflexion to this scenario: Trial 1: Agent queries 'sales data' without date filter → wrong results. 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 Reflexion (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 Reflexion happens in the code.

1def reflexion_agent(task: str, max_trials: int = 3) -> str:  # define a reusable function2    reflections = []  # key line for Reflexion3
4    for trial in range(max_trials):5        context = ""6        if reflections:  # key line for Reflexion7            context = "Lessons from past attempts:\n" + "\n".join(reflections)  # key line for Reflexion8
9        result = agent_attempt(task, context)10        score = evaluate(result, task)11
12        if score >= 0.9:13            return result  # return the result14
15        reflection = generate_reflection(task, result, score)  # key line for Reflexion16        reflections.append(reflection)  # key line for Reflexion17        store_in_memory(task, reflection)  # key line for Reflexion18
19    return result  # return the result

• Vague reflections — 'it didn't work' teaches nothing
• Too many trials — wasted cost after 3 failures
• Not storing reflections for future similar tasks
• Same reflection repeated across trials — no learning
• No evaluator — agent doesn't know it failed

Quick recap — the most important points from this module.