Evidence-first debugging separates observed facts from guesses. It makes agent work easier to audit because every recommendation starts from something that can be inspected, repeated, or compared.
The first move in a debugging run should be read-only. Capture the current error, command output, route response, file state, screenshot, or test result before changing code. If the agent changes state too early, it may hide the original cause and make the fix harder to trust.
Evidence should be specific. Instead of saying a page is broken, record which URL was opened, which viewport was used, what status code returned, what visual issue appeared, and which file or component likely owns it. That turns debugging from opinion into a trace.
A good debugging note should keep facts, interpretation, and action separate. Facts are observed directly. Interpretation explains what the facts likely mean. Action is the smallest change that should improve the state. Verification then proves whether the action worked.
This pattern is especially useful for AI agents because it prevents confident but unsupported edits. The agent should be able to point to the evidence that justified the change and the verification that confirmed it.
When the run is complete, the final report should include the original symptom, the evidence collected, the change made, and the verification result. If verification was skipped, the note should say why instead of implying certainty.