Analysis of ~400,000 Claude Code sessions shows that humans make planning decisions while AI handles execution, with higher expertise leading to more work per instruction. Debugging time nearly halved over seven months, and task value rose by about 25% on average.
Anthropic released a research report analyzing ~400,000 Claude Code sessions from ~235,000 users between October 2025 and April 2026, using a privacy-preserving framework. The study examined human-AI collaboration patterns, task success rates, and the impact of domain expertise in agentic coding.
As agentic coding tools surge in adoption, questions arise about whether non-coders can direct agents for complex technical work and how these tools will affect knowledge work labor markets. This study provides early signals based on real-world usage data.
The findings suggest that domain expertise, not coding proficiency, amplifies effective use of agentic tools. While domain experts achieve higher success rates, the gap with intermediate users is modest, indicating that moderate domain knowledge suffices for near-expert performance. The decline in debugging time and rise in task value imply AI agents will increasingly handle more complex and valuable work.
Some users express reluctance about agentic coding tools analyzing their data, insisting on being asked for consent via a survey link. This highlights concerns over data privacy and user autonomy. The comment underscores that privacy is a more critical issue than the tool's convenience.