A benchmark of frontier AI agents on production incident response: twenty incidents across three real systems, with per-trial cost and full agent trajectories.