Self-Correction
An agent noticing it went wrong and fixing the mistake on its own, without a human stepping in.
Think of it like
A cook tasting the sauce, realizing it’s over-salted, and adding cream to pull it back — mid-service, no manager needed.
Example
The agent runs its code, sees a stack trace, reads the error, patches the offending line, and reruns until it passes.
How it actually works
Self-correction is reflection turned into action: detect the error, diagnose it, retry. It works best when the environment gives honest feedback — failing tests, API errors, a search that returns nothing — because that grounds the fix in reality rather than the model’s guesses. Without such signals it can “correct” a right answer into a wrong one, so the quality of the error signal is everything.
For product teams
What lets agents recover from stumbles instead of confidently shipping the stumble.
For engineers
The act of detecting a failure from feedback and revising the action, ideally driven by external error signals.
Related
- Reflection — The critique step that precedes the fix.
- Observation — The error signals that make correction trustworthy.
- Replanning — When correction means redoing the whole plan.
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