The framing does a lot of work: "leads" means Claude drove the bulk of a task while a human stayed in the loop, and Anthropic is careful to say the model isn't running autonomously — about 90% of R&D now happens in collaboration with it. Still, the trajectory is the story. Zero in February, a quarter by August, and a fleet of some 30,000 agents pushing research and engineering forward. The more interesting half is what they bolted on to watch it. Anthropic says it monitors for agent misbehavior and has committed to embedding external third-party evaluators inside the company to keep an eye on its safety work. Whether that's real accountability or a preview of the governance arguments everyone else will be having in a year, it's worth reading before the rest of the field starts quoting these numbers back at you.