A coding agent should be judged after the verification loop

Model comparison note: judge agents by tests, scoped diffs, and reviewability rather than first-answer polish.

The first answer is often the wrong place to judge a coding model. A better check is what happens after it touches real files: Can it keep the patch small? Can it run the relevant test? Can it explain the failure without hiding it? Can it review its own diff like a maintainer would? That is why two models can both feel “smart” in chat but behave very differently inside a repo.