Anthropic says its analysis of roughly 400,000 Claude Code sessions suggests that success with AI coding tools depends more on domain expertise than on general programming skill.
The company’s latest economic research, published Monday, draws on privacy-preserving analysis of interactive sessions from about 235,000 people between October 2025 and April 2026. Anthropic said the study examines what users are doing with Claude Code, how much autonomy the assistant has inside a session, and which kinds of users are most likely to get useful results.
The broad takeaway is that people usually steer the problem while Claude handles much of the implementation. Anthropic found that users make most of the planning decisions, while Claude makes most of the execution decisions. On average, people were responsible for about 70% of planning, such as deciding what to build or what counts as done. Claude, meanwhile, made about 80% of the execution choices, including which files to edit, what commands to run, and what code to write.
Anthropic categorized sessions into nine kinds of work. About 56% involved writing, fixing, testing, or coordinating code. Another 17% were about operating software, such as deploying or monitoring systems. Planning and exploration made up 14% of sessions, while 13% centered on analysis or prose-based communication rather than coding itself.
The company said usage shifted over the seven months studied. Debugging accounted for a smaller share of sessions by the end of the period, falling by nearly half. At the same time, people increasingly used Claude Code for more end-to-end tasks, including running and deploying code, analyzing data, and writing non-code documents.
Anthropic also said the value of the typical task rose over time. Based on a comparison with freelance job postings, it estimated that the average task became about 25% more valuable across the period it studied.
A central finding of the report is that task-specific expertise appears to matter more than whether someone is a trained software engineer. Anthropic said its classifier looked for signs such as how precisely a user framed instructions, what they asked Claude to verify, and whether the user corrected the model or vice versa.
The company emphasized that expertise in this context is not the same as a job title. A senior engineer can still be a novice in a language they have never used, while a non-programmer can be highly skilled at directing Claude if they deeply understand the task at hand.
That pattern showed up in the results. In novice sessions, each prompt triggered about five Claude actions and roughly 600 words of output. In expert sessions, one prompt led to about 12 actions and about 3,200 words of output. Anthropic said the gap held across different work types and value levels.
The company also said more expertise was linked to a higher chance of ending a session successfully. The difference between intermediate and expert users was described as modest, but the trend remained clear: users who understood the problem better were more effective at getting Claude to complete it.
Anthropic framed the findings as an early indicator of how AI agents may affect work beyond software development. The company suggested that coding agents are not simply replacing human judgment. Instead, they may be reducing the amount of implementation work people need to do while rewarding those who can define problems clearly and verify the results.
The report comes as agentic coding tools gain traction more broadly. Anthropic said the share of GitHub projects with coding agent activity has more than doubled since late 2025, and that Claude Code users now spend an average of 20 hours a week using the tool.
For now, the company said the evidence points to a stable division of labor: people decide what they want, and the AI increasingly determines how to get there.