Reflection
The agent pausing to critique its own work and decide what to fix before moving on.
Think of it like
Rereading your own email and thinking “that middle paragraph is a mess” before you hit send.
Example
After drafting code, the agent reviews it, notices an unhandled edge case, and revises before running the tests.
How it actually works
Reflection adds a step where the model evaluates its last output against the goal and produces feedback it then acts on. It reliably catches some classes of mistakes and improves quality — but it has a ceiling, because the same model doing the work is grading it, and it can be blind to its own blind spots or talk itself into false confidence. External signals like tests or tool results make reflection much sharper.
For product teams
A cheap quality bump for agent output, with a real limit — a model can’t always see its own errors.
For engineers
A step where the model critiques its prior output and generates corrective feedback fed into the next iteration.
Related
- Self-Correction — What reflection leads to once a flaw is spotted.
- Agent Loop — The loop reflection is one step within.
- Tool Result — Grounding the critique in real feedback.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome