LLM reviewers are useful, but some PR checks should stay deterministic

Short summary

AI-generated PRs need both LLM judgment and deterministic CI gates. The author proposes mechanical checks—scope boundaries, workflow permissions, agent control files, test evidence, and reviewer routing—that should stay automated and repeatable. Start with warnings before promoting findings to blocks.

• •AI agent PRs need a mix of judgment-based human review and deterministic automated checks
• •Specific mechanical checks work best: PR scope, workflow permissions, control plane files, test evidence
• •Start with warnings and observe which findings are useful before enabling blocks

Generated with AI, which can make mistakes.