Turn the brief into observable behaviour.
“Build a contact form” leaves several questions open. Should empty messages be rejected? What happens when delivery fails? Can a visitor use it with a keyboard? Acceptance criteria make those questions part of the job.
Tests should exercise the application’s behaviour, not simply confirm that the implementation repeats its own assumptions. AI reviewers add judgment about the brief and changed code, while executable tests provide a different kind of evidence.
Run relevant checks and widen when needed.
Vraido’s test planning can map changed files to affected tests and use the project’s known test commands. A narrow change may justify a focused run. Shared configuration, unknown dependencies or an unrecognized test setup calls for broader checking.
The changed-file list comes from version control. It should not depend on an agent remembering to mention every file it edited. This is another coordination job that software can perform directly.
Reuse a result only while its inputs match.
Passing evidence belongs to a test group, its relevant inputs, environment and requirements. Reuse avoids repeating unchanged work, but a cached pass is not interchangeable with a fresh pass after its assumptions change.
Skipped, failed, reused and not-run checks mean different things. Keeping those distinctions visible makes it easier to judge the result without opening every command log.
Check the combined result.
Two branches can pass independently and fail when brought together. Integration checks need to cover the candidate containing the combined changes. A coder finishing a branch is an intermediate event, not the final definition of completion.
The same principle applies to publishing: a local preview, merged code and a deployed service are different states. The evidence should identify the environment you are looking at.
Evaluate the AI workflow too.
Model evaluations need representative work, consistent starting conditions and defined acceptance criteria. Compare planning, coding and review roles separately, then measure the entire workflow.
A few successful toy tasks do not establish production reliability. Use the results to improve the setup, and keep limitations visible when sample sizes or cost measurements are incomplete.