A bug report is not a root cause. A stack trace is not a fix plan. The first useful debugging act is to produce a counterexample: a minimal case that makes the system fail in the way the hypothesis predicts.
AI agents skip this step when they are allowed to. They read the symptom, jump to the most familiar cause, patch the nearby code, and explain the patch well. Sometimes that works. The rest of the time it leaves the actual cause alive and adds defensive code around a misunderstanding.
Evidence before hypotheses
The order matters. Read the failure output, the logs, and the route, handler, test, or page that produced it before forming a theory. Then create two or three falsifiable hypotheses. Falsifiable means the hypothesis predicts something you can check. If the check does not fail as predicted, the hypothesis is wrong or incomplete.
This discipline stops the common agent failure where a plausible explanation becomes a magnet. Once the agent has told a coherent story, it starts interpreting every file through that story. A counterexample breaks the spell.
The reproduction is the hinge
A useful reproduction is small enough to run quickly and specific enough to prove the failing behavior. It might be a unit test, an integration test, a curl command, a browser trace, or a database query. The form matters less than the contract: it fails before the fix and passes after the root cause is removed.
Without that hinge, you are not debugging. You are editing near a symptom.
Do not code around the environment
Sometimes the root cause is not code. A service is down. An env var is missing. A local database is stale. A permission is absent. The disciplined answer is to classify it and stop, not to add fallback logic that hides the real operational failure.
A good debugging prompt makes that classification explicit: in scope, sibling change, environment, or real same-track failure. The agent earns the right to edit only after it has proven which bucket the failure belongs to.
Series linkage
Part 6 of 10 in Prompt Library to Operating System. Review catches design and branch failures; debugging handles the failures that escaped into behavior.
About AmanERP
AmanERP's engineering loop favors root-cause fixes over symptom patches because business software cannot be calm when failure modes are merely hidden.