An agent gate inserts AI review into delivery, but separates probabilistic advice from deterministic enforcement. Secret scanners and other reliable checks may block; models should normally warn. section 5.3.2
A disciplined agent gate follows these rules:
- Run deterministic checks such as gitleaks before accepting a commit. section 5.3.2
- Ask the model about questions linters cannot reliably answer, including error-handling philosophy, naming, conventions, and test coverage. section 5.3.2
- Parse a structured verdict instead of grepping free-form output, and keep model findings advisory to avoid habituating developers to bypasses. section 5.3.2
