Fix bugs with an AI coding agent by requiring a reproducible failure before accepting a repair. Provide the observed behavior, expected behavior, environment, logs, recent changes, and the smallest known reproduction. Ask the agent to explain its hypothesis and evidence before editing code.
Have the agent trace the failing path and identify where actual state first differs from expected state. Convert the reproduction into a focused automated test when practical. Evaluate one or two plausible causes, then choose the smallest patch that addresses the cause rather than hiding the symptom. Run the focused test, nearby tests, and the project’s standard checks. Inspect error handling and neighboring call sites for the same assumption. If the issue involves data, migrations, concurrency, permissions, or external services, add a rollback or containment plan before deployment.
Reject fixes that only silence an exception, weaken validation, delete a failing test, or add retries without explaining the failure. The completion report should include root cause, changed files, tests run, results, and anything that could not be reproduced.
Start the agent in an isolated branch with no production credentials. A good bug-fixing loop leaves behind a test that would have caught the defect and a short explanation another maintainer can verify. That is stronger evidence than a patch that merely appears to work once.
