Self-Correction

Self-correction is the AI's ability to detect that its own output is wrong (via tests, schema validation, or a critic model) and retry — without human intervention.

Self-correction is reflection + action: the AI not only notices a problem in its output but loops to fix it. Common forms: code that won't compile triggers a fix retry, JSON that fails schema triggers a re-emit, an answer flagged as low-confidence triggers an additional retrieval round.

Example

AI Engineer writes a function. Tests fail. The error message goes back into the model with the diff. AI rewrites. Tests pass. PR opened — without a human knowing the first attempt failed.

How OpenLabor uses it

OpenLabor agents auto-retry on validation failures up to a bounded count, then escalate.

Can self-correction loop forever?

Without bounds, yes. Always cap retries; escalate when the cap hits.

Related: reflection, eval, agent-framework.

AI Labor Glossary