Stop revisions from piling up in the queue

Each review request should lead to one bounded next pass. Group related comments, mark which are blockers and which are optional, and keep the original acceptance criteria. If review turns up separate work, open a new task for it instead of hanging it off the current change, unless release safety depends on it.

Ask the agent to explain how it dealt with each blocking comment and which tests changed. The reviewer should be comparing the new diff with the version they last reviewed, without starting again from the repository base every time. Force-pushing can be fine in a controlled workflow, but you need enough history or review tooling to see what changed between rounds.

I'd set a revision limit as a prompt to reassess, not an automatic rejection. When a change keeps going round, it's often because the task is ambiguous, nobody owns the design, or the agent is working past the point where its context window is still useful. Stop, rewrite the task and start a clean implementation if that's cheaper than untangling the branch.

Watch for patches that just accumulate. An agent might keep its earlier work and wrap a new conditional around each comment until the tests pass. That can satisfy every request and still leave something hard to reason about. After a substantial revision, review the whole final design again and ask whether any path that's been superseded is still in there.

Treat merge conflict resolution as new work. A mechanical resolution can change behaviour when both branches touched the same contract. Rerun the affected tests and have a reviewer look at the combined result. An agent that wrote one side shouldn't be quietly deciding what the other side meant.

Close rejected branches and write down why. They shouldn't show as active on the task board, and later agents shouldn't find them and treat them as usable prior work. Keep any useful diagnosis in the issue, then clean up the disposable worktrees and branches the way your repository policy says to.