The judge gate: why a passing validator isn't a finished feature

Autonomous coding agents can stop too early when tests and builds pass, even if the code is still stubbed or incomplete.

The piece compares goal/loop patterns such as OpenAI Codex, the Ralph loop, and Claude Code’s goalkeeper approach.

It proposes a separate judge agent that reviews the full diff in fresh context against a strict checklist before declaring work done.

In a real Claude Code plugin workflow, the judge rejected a benchmark test that relied on a sentinel runtime value even though the validator passed.

The takeaway: green tests can hide false positives, so an independent code-review gate can prevent flawed work from shipping.