Reflection
Reflection is asking the model to review its own work before it ships. It helps more than people expect and less than the papers suggest, and everything depends on whether the critic has fixed criteria or is free to say looks good.
How it works
- The agent produces a first attempt.
- A critic pass evaluates it against written criteria, not against taste.
- Failures come back as a specific list, not as a vibe.
- The agent revises once, addressing only the listed failures.
- If the criteria still fail, it stops and hands the work to a person rather than looping.
Optional: load a real model
See it work
When it pays, and when it does not
| Use it when | Skip it when |
|---|---|
| The output has criteria you can write down | You cannot say what good looks like, in which case reflection just adds cost |
| A second pass is cheaper than a human review | The task is trivial and the first attempt is nearly always right |
| You can constrain the critic to a checklist | The critic is the same prompt as the writer, which mostly produces agreement |
| One revision round is enough | You are tempted by five rounds, which drift rather than improve |
How it fails
Self-congratulation
A critic with no criteria says the draft is strong and adds nothing. Fixed criteria or no reflection.
Drift across rounds
Each revision moves further from the source facts. Cap at one round and re-check facts after it.
Cost doubling for nothing
Reflection doubles or triples token cost. Measure whether the second pass changes outcomes before keeping it.
What it costs
| Latency | At least two model calls instead of one, often three. |
|---|---|
| Tokens | Two to three times a single pass, which is the main reason to check that it earns its place. |
| Complexity | One more failure point, and one more place where a prompt change silently alters behaviour. |
Deterministic critics beat model critics. Where you can express a criterion as code, do, and save the model for the rewrite.
Related: Multiple agents · Verification · Testing agents · all patterns · agent jobs
Related: Multiple agents · Verification · Testing agents · all patterns · agent jobs
Free from AI School - no signup, everything runs in your browser.