To loop an AI agent until tests pass, give it a clear test command, a failure condition, and a stopping rule. The agent should run the tests, read the failure, change the code, rerun the tests, and stop when the test passes or when the failure requires human judgment.
The stopping rule matters. Without one, the agent may keep making shallow fixes, hide the real issue, or spend too much model usage on an unclear problem. A good loop should include the target test, maximum attempts, files in scope, and what to do if the agent is stuck.
Verdent's stronger workflow is loop plus planning and review. Plan the task first, let the agent generate and test changes, keep work isolated, then review the final diff. The goal is not infinite self-correction. The goal is a bounded generate-test-fix cycle that produces code a human can trust.
