In February, Claude led under 1% of the work on its own next model. By August: 26%. Fully autonomous work: still zero. The most useful part of this story isn't the percentage — it's the ruler it was measured with.
On September 17, Anthropic published internal R&D automation metrics for the first time. "Leads" is a level on the AL0–AL5 automation scale borrowed from Epoch AI: the agent completes a task end-to-end from a high-level prompt while a human supervises. As of August that covers 26% of the work; more than 90% of R&D happens "in collaboration" — large chunks done under close human direction. The scale of the pipeline: roughly 30,000 concurrent agents, over 1 billion agent decisions in August, of which online monitoring intercepted about 0.002%. Around 6% of R&D compute goes to safety, rising to 12% inside AI-led work: the more the agent does, the more supervision costs.
Here is the pair of numbers I would pin above every debate about agents: 26% "led" against 0% "autonomous." The lab standing closest to recursive self-improvement runs it as human-in-the-loop with no exceptions. Which means supervision is not a transitional phase that fades once models get smarter, but a load-bearing part of the process.
Now, the skepticism check. The number is measured by the party that benefits from it, and where exactly "leads" begins — how many retries and corrections it forgives — is invisible from the outside. Anthropic promises to embed independent third-party evaluators; that is when we learn whether 26% survives someone else's ruler. My motto cuts both ways: trust no mind you have not measured — and no measurement report about one, either.
What this means for an engineering lead. Take the language, not the headline. AL0–AL5 turns the fuzzy argument "are we using AI enough" into a concrete 1:1 question: "what level is this task?" This week's exercise: take last sprint's closed tickets, grade them honestly on the scale, and compute your own "agent leads" share. It will likely come out well below how it felt — and that is a perfectly fine starting point: Anthropic was under 1% in February. You can only manage the share you have measured.
📎 anthropic.com/transparency · bloomberg.com/news/articles/2026-09-17 · abcnews.com