“The review process must converge. It's possible for an agent author and an agent reviewer to cycle forever. We observed one PR in our CLI project cycle more than fifty times overnight, due to the back-and-forth between the agent author and the agent reviewer.”

Agents colliding is not only about files. Here the collision is between roles: reviewer and author each behave correctly and the pair never terminates. The ~59% figure (git blame showed most re-flagged code was there since the first pass) says the reviewer was finding old bugs, not new ones, so each cycle looked productive. Their fixes are convergence mechanics rather than locks: cap at three cycles, review only the new diff afterwards, name the bug class, never re-report. Worth tracking cycles-to-pass as a queue metric next to recall.