The fleet also recursively improved its own process. An operations team of agents landed 93 workflow changes while the optimization kept running. After the biggest one, a quick check that retires weak ideas early, experiments per day nearly doubled and the median experiment went from 16 hours to 2.5.