Agents Marking Their Own Work as Done Is Not Verification, Argues Test-Loop Guide

The article argues that when the same agent writes and approves its output, a 'looks done' judgment reviews nothing and merely confirms the agent's existing intent. Its proposed fix separates roles: an orchestrator drives the process, a builder fixes gaps, a critic script executes real checks and produces a score, and a separate reviewer in a fresh context signs off. Checks are declared up front with commands and weights, and rounds are capped so non-converging loops stop. A fail-closed security gate blocks delivery on secrets or high-severity findings even when tests pass. The piece also warns of pitfalls such as pipes masking exit codes, timeouts silently scoring zero, and loose or regenerated baselines that make failures look like passes.