Reflection

Reflection is a pattern where an AI reviews its own work — checking for errors, gaps, or hallucinations — and revises before producing the final output.

A single forward pass is what the model 'thinks' on first read. A reflection pass is what it 'thinks' after re-reading. Empirically, reflection catches a meaningful fraction of errors — especially in code, math, and synthesis tasks. The cost is more tokens and more latency; the benefit is fewer wrong answers.

Example

AI Engineer drafts a code change, runs tests, reads the diff against the requirements doc, notices a missed edge case, patches, then opens the PR. The reflection pass catches the bug the first pass missed.

How OpenLabor uses it

OpenLabor adds reflection passes on high-stakes missions automatically.

Is reflection the same as chain-of-thought?

Related but separate. CoT is reasoning before answering; reflection is reviewing after answering.

Related: chain-of-thought, self-correction, eval.

AI Labor Glossary