Odaily reported that Anthropic has released a large-scale report on agentic coding based on Claude Code usage. The report, covering 400,000 sessions, says that when Claude Code assists with programming work, coding skill itself becomes less decisive, while management ability and domain knowledge become key factors shaping the result.
Non-programmers narrow the success gap
According to the report, Claude Code users are active for an average of 20 hours per week. Across the ten main occupations studied, the success-rate gap between non-programmers using agents to write code and professional software engineers has narrowed to within 7%. Management-role users recorded a success rate slightly higher than that of professional programmers.
The report also describes how decisions are divided between users and the agent. Users lead about 70% of planning-direction decisions, while the agent handles about 80% of concrete execution decisions. Expert-level sessions can trigger 12 actions and produce 3,200 words in a single session, more than five times the level seen in beginner sessions.
Debugging declines as deployment and analysis grow
User experience level affects how sessions continue after errors. Beginner sessions have a 19% abandonment rate after encountering error messages. For intermediate and expert users, the abandonment rate when facing problems is between 5% and 7%.
Over a seven-month period, the share of debugging sessions used to fix broken code fell from 33% to 19%. At the same time, the share of tasks involving direct deployment and running, data analysis, and writing non-code documentation doubled. The report says the average estimated value of tasks increased by about 25%, reflecting a shift in how Claude Code is used across different types of work.