Le cliquet de mesure
Before asking an agent to speed anything up, you build a bench that reproduces the slowness with a stable counter, you prove that counter tracks the time the user actually feels, then you turn it into a continuous integration threshold that can only go down. This is the method the claude.ai team published on 23 September 2026: 13 measurements across 4 journeys, 3.1 times faster on average in two weeks, more than 3,000 changes without a customer incident or rollback.