Anthropic just dropped a research report that reads like a performance review for its AI coding agent. The verdict: Claude Code is great at doing things, but humans are still the ones deciding what to do.
The report, titled “Agentic coding and persistent returns to expertise,” analyzed roughly 400,000 Claude Code sessions from about 235,000 users between October 2025 and April 2026. It’s part of Anthropic’s broader Economic Index series, which has been quietly building one of the more rigorous datasets on how people actually use AI tools in the wild.
The division of labor is stark
Users are responsible for approximately 70% of planning decisions, essentially telling Claude what needs to happen. Claude, in turn, handles about 80% of execution decisions, figuring out how to get it done.
Domain expertise, the kind of knowledge that goes beyond just knowing how to code, makes a massive difference. Expert users coax Claude into generating over twice as many actions per prompt and five times more output compared to novices.
