Anthropic's AI now leads a quarter of its own R&D. The way it's supervised is a template for your own AI agents

Anthropic disclosed this week that Claude now leads 26% of the work that builds the next version of Claude, up from close to none in February — and that roughly 30,000 AI agents were doing research and engineering work at the company as of August. The number that matters more for a construction firm isn't the 26%. It's the oversight structure Anthropic had to build to let that number climb without losing control of the work, because that structure is a usable template for any GC now handing real decisions to an AI agent of its own.

What did Anthropic actually disclose?

In a report from its internal Anthropic Institute, the company said the share of its research and development that Claude "leads" — meaning it does a task largely on its own, with a human reviewing the output rather than every step — went from essentially zero in February to 26% by August. Separately, Claude now "collaborates" on more than 90% of Anthropic's R&D, meaning it does large chunks of work under close human direction. Running that much autonomous activity required a formal oversight system: continuous monitoring for agent misbehavior, external third-party evaluators being embedded inside the company to audit safety practices, and a stated principle that the ability to slow down or stop the work is non-negotiable.

Why does an AI-lab staffing detail matter to a GC?

Because the underlying problem is the same one showing up on your projects right now, just at a different scale. This week alone, xAI disclosed a procurement bot that flagged $100,000 in vendor waste on its own, and OpenAI's Agents API is already being built into construction software platforms. Every one of those tools forces the same question Anthropic had to answer for its own agents: how much of this task does the agent lead, versus collaborate on with a human checking each step? Most construction firms adopting an AI agent skip that question entirely — they turn the tool on and see what happens. Anthropic's disclosure is, in effect, a published answer key for how a serious organization draws that line.

What does "lead" vs. "collaborate" look like on a jobsite?

The construction equivalent isn't abstract — it maps onto an authority matrix almost every GC already runs for change orders and RFIs, just applied to an AI agent instead of a project engineer.

Anthropic's own numbers show it hasn't moved its most consequential work into that third row, even with a 30,000-agent research operation. That's the tell: a lab racing to build smarter models is still keeping human sign-off on anything with real consequences, and moving agents to "leads" status only on work a human can check after the fact without much damage if it's wrong.

What should a GC or sub do with this?

Before turning on an AI agent for procurement, scheduling, submittal review, or anything else, write down — in one sentence per task — which of the three rows above it sits in today, and who checks the output and how often. Don't let a vendor's demo answer that question for you; ask what happens when the agent is wrong, and whether that failure surfaces before or after money moves. If a vendor can't describe their own version of Anthropic's monitoring-plus-stop-button setup, that's a gap worth pricing into the decision, not a detail to skip.

This week's xAI procurement-bot story covered an agent finding real savings on its own — read together, these two pieces are the "what an agent can do" and "how much rope to give it" halves of the same decision.

Forward this to whoever's about to approve the first AI agent with sign-off authority on your project.

Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.