Agentic Coding and the Persistent Returns of Domain Expertise

Anthropic analyzed approximately 400,000 Claude Code sessions from October 2025 to April 2026 to investigate the actual usage patterns of agentic coding. The analysis revealed a clear division of labor: users handle Planning, deciding what to do, while Claude handles Execution, deciding how to implement it.

The success of agentic coding was found to be more influenced by Domain Expertise than by coding proficiency. The richer the domain expertise, the more work Claude could accomplish from a single instruction, and the faster recovery was possible when errors occurred. However, the gap in success rates between highly skilled experts and intermediate-level users was not large, suggesting that anyone with a certain level of domain knowledge can use the tool effectively.

Usage patterns are also changing. Over the past 7 months, time spent on debugging decreased by nearly half, while the share of end-to-end agentic use—such as code deployment, data analysis, and non-code documentation—increased. Additionally, based on comparisons with the freelance work market, the value of typical tasks through agentic use rose by an average of 25%.