How to Debug AI Coding Agents When They Change the Wrong Thing

AI coding agents can claim a task is "fixed" without providing evidence of what was inspected, changed, or verified. A practical debugging method addresses this traceability gap by recording the agent's tool requests, function results, applied patches, and verification checks into a run log. The process is demonstrated by fixing a UI bug, where the agent's actions like file inspection, applying a CSS patch, and capturing screenshots are logged for review. This detailed log provides the necessary evidence to confirm the agent correctly understood the request, targeted the right code, and successfully resolved the problem.