Anthropic says coding agents are starting to shift technical work away from pure programming skill and toward domain expertise.

In a new report based on about 400,000 Claude Code sessions from roughly 235,000 users between October 2025 and April 2026, the company found that people are increasingly using AI coding agents for more complex and higher value work. 

The data covers Claude Code usage across its command line interface, Claude.ai, and the Claude Code desktop app.

The report shows a clear split in how people work with coding agents. Users made about 70% of planning decisions, such as what to build or what should count as done, while Claude handled about 80% of execution decisions, including which files to change, what code to write, and which commands to run.

Anthropic said that means users are still setting direction, while the agent is increasingly deciding how to implement the work.

About 56% of Claude Code sessions involved writing, fixing, testing, or orchestrating code. Another 17% involved operating software, while 14% focused on planning or understanding systems. 

Around 13% involved analysis or prose based work, showing that Claude Code is being used beyond traditional software development.

The company also found that the type of work being done has changed quickly. The share of sessions spent fixing broken code fell from 33% in October to 19% in April. At the same time, operating software rose from 14% to 21%, while writing and data analysis roughly doubled from about 10% to 20%.

Anthropic estimated that the average value of Claude Code sessions increased 27% over the period, with building, operating, and fixing tasks all becoming more valuable.

The strongest signal in the report was user expertise. Novice sessions reached Anthropic’s strictest measure of verified success 15% of the time, while intermediate and expert sessions reached verified success between 28% and 33% of the time.

The gap was especially visible when sessions ran into trouble. Novice users were more likely to abandon failed sessions, while more experienced users were better able to recover, correct the agent, and steer the work toward completion.

Anthropic said the data suggests that coding agents are not replacing domain expertise. Instead, they appear to amplify it. Users who understand the problem they are solving can get more out of the agent, even if they are not trained programmers.

That pattern also showed up across occupations. Software related users reached verified success in about 30% of sessions overall, compared with 26% for users from other professions. Among sessions that actually produced code, the gap was 34% versus 29%.