Coding skills are becoming irrelevant. That's the headline takeaway from Anthropic's massive new study of 400,000 sessions of Claude Code, its AI-powered coding agent. The real differentiator? Domain knowledge and the ability to manage the agent.
According to the report, the average Claude Code user spends a whopping 20 hours per week actively using the tool. Across ten major professions, non-programmers using the agent to write code now succeed at rates within 7% of professional software engineers. In fact, managerial roles slightly outperformed the coders themselves.
The study reveals a striking division of labor: humans make about 70% of high-level planning decisions, while the agent handles roughly 80% of the execution-level choices. Business experts who know how to break down and clearly delegate tasks get far more out of the model. An expert session triggers an average of 12 actions and produces 3,200 words of output — over five times that of a novice session. Novices, on the other hand, are quick to give up when they hit errors: the "frustration abandonment rate" for beginners sits at 19%. For intermediate and expert users, that number drops to just 5–7%.
The study argues that with just intermediate knowledge in a specific domain, anyone can use an AI agent to leapfrog the programming barrier and dramatically boost personal productivity.
Developers are also shifting how they use agents. Over the past seven months, debugging sessions — fixing broken code — have dropped from 33% of all sessions to just 19%. Meanwhile, sessions dedicated to deploying code, analyzing data, and writing non-code documentation have doubled. The average estimated value of tasks users delegate to the agent has risen by about 25% in the same period.