A team that adopts AI coding agents gets more changes, larger changes, and a new kind of change: edits to the agent itself. Review queues and CI capacity become the bottleneck, and the slow, shared end-to-end test environments fill up first. This site follows one change from diff to production, through code change, commit, pull request, review and CI/CD, and weighs the approaches at each stage with the drawback of each.
Before agents, the slow step was writing the change, and review and CI were sized for that pace. With agents, a developer can open several pull requests in the time one took before. Nothing downstream got faster on its own. People skim reviews or the queue stalls, and the few end-to-end test environments fill up with runs that a lint would have stopped.
After each edit: fast deterministic checks. Before the agent stops: the affected tests. Before a push: the guard. What each hook can and cannot do, and why only the server is the gate.
Review on the developer machine before the pull request exists, then let CI verify that work by hash instead of repeating it. The laptop pre-pays; it is never trusted.
Sort changes by what they can break, not by who wrote them. Merge queues, test selection and scheduling for scarce end-to-end environments. For a red build: triage, then a deterministic verify, then fix.
Reviewing a skill file like code, and building an eval that is fast, parallel and still honest: freeze the world instead of replaying the path, pick the smoke set from the diff, grade in three layers, and grow the test set from shadow runs.