Anthropic published a research paper yesterday based on an analysis of roughly 400,000 Claude Code sessions. The title: “Agentic coding and persistent returns to expertise.” It sounds academic, but the findings are remarkably practical for anyone working with AI coding tools.
Key findings
The study covers October 2025 through April 2026 and paints a clear picture of how people and Claude collaborate:
People plan, Claude executes. In a typical session, users make most of the planning decisions — what to do. Claude handles the execution — how to do it. Sounds like a simple division of labor, but the data shows: the more experienced the user, the more the balance shifts toward Claude doing heavier lifting.
Expertise is the multiplier. This is the real headline: the more domain expertise someone brings, the more work Claude does per instruction. It makes sense — precise prompts yield better results.
Non-developers are surprisingly successful. On coding tasks, people from every major occupation succeed at nearly the same rate as software engineers, as long as they bring domain expertise. A product manager who knows their domain can work with Claude Code just as effectively as a senior developer.
Growth confirmed
The study also delivers hard usage numbers: the share of GitHub projects with coding agent activity has more than doubled since late 2025. Claude Code isn’t an experiment anymore — it’s infrastructure.
My take
What strikes me most about this study: it confirms something I’ve experienced firsthand. As a product manager with a development background, Claude Code was a productivity booster from day one. Not because I write perfect code, but because I know what I want — and can articulate it precisely.
Anthropic’s message is clear: agentic coding doesn’t replace expertise. It amplifies it.
Sources: Anthropic Research