AI is helping build AI

AI is helping build AI.

Anthropic reported on September 17 that Claude now leads 26% of its internal model research and development work.

The figure was 1% in March.

“Leads” means completing most of a task from a high-level prompt while a person supervises.

Claude also collaborates on more than 90% of Anthropic’s research work. No measured work operates fully autonomously.

This distinction matters.

The story is not autonomous AI building its own successor.

The shift is AI taking a larger role in the research cycle responsible for creating future models.

Anthropic also reported:

• Around 30,000 agents performing research and engineering work
• More than one billion agent decisions screened during August
• Approximately one in 47,000 decisions blocked before execution

For technology leaders, adoption metrics are no longer enough.

Organizations should also measure:

• Work led versus supported by AI
• Decisions executed or rejected
• Human-review requirements
• Quality and safety outcomes
• Business value created
• Accountability for agent-produced work

My takeawayas AI agents perform more complex work, governance must measure both contribution and control.

Does your AI operating model track how much work agents lead, what they decide, and where people intervene?