Agent self-correction uses feedback from tests, tools, observations, critics, or explicit checks to identify a mismatch between the current result and the goal. The agent then diagnoses the issue, chooses a corrective action, and verifies the updated state.
Self-correction is not guaranteed to improve a result. An agent can misread feedback, repeat the same error, damage working parts, or stop after a superficial fix, so important tasks still need bounded retries and independent verification.
ELI5
Agent self-correction happens when an AI agent uses outside evidence to notice that its attempt failed, changes its approach, and tries again. The evidence should come from something checkable rather than the agent simply deciding that its own answer looks better.
For example, an agent may edit a program, run its tests, see a specific failure, and make a focused repair before rerunning the test. Retries should stop when the same failure repeats, the evidence is unclear, or the next change would require different permission.
What evidence can trigger agent self-correction?
Compiler errors, failed tests, logs, screenshots, tool responses, environment state, human feedback, critics, and explicit requirement checks can trigger it.
Why does self-correction still need verification?
The agent may misunderstand the failure, introduce another defect, optimize the wrong measure, or claim success after only a partial repair.
