Anthropic: Agentic Coding Study — Users Make 70% of Planning Decisions, Claude 80% of Execution Decisions

Anthropic published an economic study of agentic coding with Claude Code that quantifies how work is divided between the human and AI agent.

How are decisions split between users and Claude?

According to the study, users make approximately 70% of planning decisions, while Claude takes approximately 80% of execution decisions. In other words, the human retains strategic direction (what and why), while the agent handles most of the operational execution (how). This division of labor provides a framework for understanding where human expertise still matters most and where the agent carries the burden.

How much does expertise affect outcomes?

The difference between experts and beginners is substantial. Expert users receive 12 actions and 3,200 words per prompt versus 5 actions and 600 words for beginners — more than double. Verified task success rates are 28–33% for expert and intermediate users versus 15% for beginners. Notably, workers outside software succeed in 26% of coding sessions, just 4 percentage points below software engineers (30%).

How is the nature of work changing with agents?

During the tracked period, the share of debugging sessions fell from 33% to 19% (October 2025 – April 2026), while the share of “software management” rose from 14% to 21%. Average task value increased approximately 25%. The trend shows a shift from error correction toward orchestration, suggesting that agents are taking over an ever-larger share of routine execution.