Normal text based analysis; which LLMs are pretty good at. So our coding agents can follow the rules during development and the review agent can follow it during review. Only static code checks are enforced in the CI pipeline; the review agent only leaves remarks. If we decide to not give a damn, we can still merge it. But its hit-rate is relatively high; so typically when it finds something, it’s valid and its recommendations for fixes are also often pretty good. This helps already iterating on a PR on your own before you pull human reviewers in.