Skip to main content

Claude Takes Over a Quarter of Anthropic's R&D Work

Anthropic has just given us a rare peek behind the curtain: its own AI, Claude, is now doing a big chunk of the company's research and development. How big? As of August, Claude led 26% of R&D tasks—a stunning leap from less than 1% just six months earlier.

The Six-Level Automation Ladder

To keep tabs on how much AI is really doing, Anthropic uses a six-level framework. At AL4, the AI takes the wheel: humans set the destination, Claude drives most of the journey, and people supervise and review. Right now, over 90% of R&D work hits that level or higher. But here's the catch—no task has reached full autonomy (AL5). Humans still hold the keys.

That human oversight isn't just a formality. Roughly 30,000 Claude-powered agents are working in parallel, assigning tasks to each other, cross-checking errors, and churning out over 1 billion decisions in August alone. Every move gets an online safety check before it runs, and anything fishy is blocked on the spot. The system also samples conversations and flags high-risk cases for manual review.

Code, Code, and More Code

The impact on coding is hard to miss. Today, more than 80% of code merged into Anthropic's repository is written by Claude. Engineers say their productivity has multiplied several times over. These agents handle the grunt work—running model experiments, writing code, analyzing data, verifying results—and even help iterate on the next generation of Claude itself. The industry calls this recursive self-improvement, and it's happening right now.

The Bigger Picture

Anthropic's experiment raises a question every tech company will soon face: how much can you hand over to AI before you lose control? For now, the answer is "a lot, but not everything." The company is betting that keeping humans in the loop—at least for now—is the safest way to let AI build AI.

Key Points

  • Claude now handles 26% of Anthropic's R&D tasks, up from under 1% in February.
  • Over 90% of R&D work reaches human-machine collaboration (AL4) or higher, but none is fully autonomous.
  • 30,000 AI agents generate 1 billion decisions per month, with safety checks and manual reviews.
  • 80%+ of merged code is Claude-generated, boosting engineer productivity.
  • AI participates in building the next Claude model, a step toward recursive self-improvement.