Reflection
Agents make wrong tool args and incomplete answers — reflection catches errors early.
Proofreading before send — catch mistakes while they are still cheap to fix.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Self-Critique Loop
Scroll inside the frame to explore · use + / − to zoom up to 200%
Draft critique revise until quality passes or max retries.
Key Takeaways
- 1.Reflection = agent reviews output before returning.
- 2.Self-critique: LLM checks quality and completeness.
- 3.Verifier: separate model judges primary output.
- 4.Reflexion: store lessons in memory for future tasks.
Real Example
Scenario
Report agent drafts a Q3 summary; critique flags a missing APAC breakdown and an uncited revenue total. Agent re-queries SQL for APAC, adds citations, then returns.
What you would do
Self-critique checks: date range ✓, all regions ✓, citations ✗. Revise triggers a second SQL call with `region='APAC'`. Cap at 2 critique rounds; use a cheaper model for critique. Monitor: critique-trigger rate and added cost per task.
Practice Task
Write a 6-item reflection checklist for customer-facing emails. Apply it mentally to a one-paragraph draft you invent — note what would fail.
Code Walkthrough
Highlighted lines show where Reflection happens in the code.
1CRITIQUE_PROMPT = "Review the draft. List: missing sections, uncited numbers, policy risks."2
3draft = "Q3 revenue was strong across regions..."4critique = llm([{"role": "user", "content": f"{CRITIQUE_PROMPT}\n\n{draft}"}]) # prompt that asks the model to review its own draft5
6if "uncited" in critique.lower():7 revised = llm([{"role": "user", "content": f"Fix these issues:\n{critique}\n\n{draft}"}])8 draft = revisedCheat Sheet
Quick recap
quick ref- •Reflect on high-stakes outputs
- •Self-critique vs verifier agent
- •Skip for simple lookups
- •Max retry after failed critique
- •Reflexion stores lessons
Common Mistakes
- ✕Skipping evaluation for Reflection before production
- ✕No logging or tracing around reflection steps
- ✕Ignoring cost and latency implications