Debugging AI coding agents requires more than just a "fixed" confirmation. The challenge lies in verifying that the agent understood the request, modified the correct code, and validated the intended outcome. Without a clear record of the agent's actions—what it inspected, changed, ran, and how it verified the fix—it's impossible to confirm if the user's original issue is truly resolved.

This is akin to software engineering's concept of traceability, connecting a user's request to the specific files, patches, command results, and verification checks performed by the agent. A practical tutorial demonstrates how to record model tool requests, actual function results, patches, checks, screenshots, and a saved run log. This evidence trail is crucial for identifying common failures, such as editing the wrong file, using an incorrect selector, or missing an edge case.

The example focuses on a user interface bug where a button overflows on a mobile screen. The repair involves a targeted CSS change. The tutorial emphasizes the importance of saved evidence, including file reads, browser checks, build outputs, DOM measurements, and before-and-after screenshots. The goal is to provide a review pattern adaptable to various coding agent errors, ensuring the agent's actions are transparent and verifiable, moving beyond a simple "fixed" claim to a summary of evidence.