Green tests are not evidence your coding agent finished the job

A coding agent can ship a green diff that is not the job you asked for.

A vague brief leaves the agent to decide what is broken, what must stay stable, and how far the change reaches. Those are product and engineering decisions. The implementation makes them for you if you leave them implicit.

The missing ingredient is rarely a better prompt. It is four loops around the agent: bound the outcome, require a plan before edits, delegate only inspectable increments, and approve evidence before merge. Green tests say the assertions still pass. They do not say the increment matched the bound. A diff shows what changed. The agent's completion report shows what it thinks changed. A mismatch is a review target.

Full write-up with a task-brief template and before-merge checklist: The Four Loops that Keep Coding-Agent Work Reliable