Nah, I saw this problem a while ago and already spec'd out the solution. First, agents need to be doing atomic commits, and second you can just have a massive merge queue with bisection, if you're using bazel you can handle ci gating on thousands of PRs with very little overhead, and when a merge batch fails you find the bad patch set in O(log(n)) time and dispatch to an agent for reconciliation. I even built a prototype, works great in benchmarks but I don't have a need for it over merge trains in gitlab yet.