The useful signal in Anthropic’s new Claude Code usage study is not “everyone becomes a software engineer.” It is narrower and more interesting: coding agents seem to reward people who understand the problem they are trying to solve.
Anthropic analyzed about 400,000 interactive Claude Code sessions from roughly 235,000 people between October 2025 and April 2026. The company’s framing matters because this is product telemetry, not a neutral labor-market survey. It also matters because the study is unusually close to the work surface: transcripts, tool activity, code changes, tests, commits, and model-based classifiers over real sessions.
The division of labor is becoming clearer. In a typical session, Anthropic says users make about 70% of the planning decisions, while Claude makes about 80% of the execution decisions. The human mostly decides what should be built, what approach counts as acceptable, and when the work is done. The agent mostly decides which files to read, what code to write, and which commands to run.
That is a different story from simple automation. The person is not disappearing from the loop. The loop is changing shape. A good user increasingly acts like a product owner, reviewer, domain expert, and test oracle, while the agent absorbs more of the implementation path between instructions.
The work is moving beyond bug fixing. Anthropic classifies about 56% of sessions as writing, fixing, testing, or orchestrating code. But over the seven-month window, debugging fell from 33% of sessions to 19%. Operating software grew from 14% to 21%. Writing and data analysis roughly doubled, from about 10% to 20%. Anthropic’s coarse estimate of task value also rose by 27% over the period.
The read is not that the estimate is a precise dollar measure. Anthropic explicitly says it is rough and based on comparisons to freelance postings. The useful point is directional: as models improved and users adapted, sessions shifted toward more end-to-end work around software, not just patching broken code.
The bottleneck moves from syntax to judgment. Sessions rated novice reached Anthropic’s strictest “verified success” measure 15% of the time, while intermediate-to-expert sessions reached 28% to 33%. The looser “at least partial success” measure was 77% for novices and 91% to 92% for intermediate-to-expert sessions.
The occupation comparison is the part to watch. Among sessions that produced code, software-related occupations reached verified success about 34% of the time, while other occupations reached about 29%. On partial success, the gap almost disappeared: 89% versus 88%. In Anthropic’s data, a coding job title mattered less than the user’s task-specific command of the problem.
That does not mean expertise is obsolete. It means the scarce skill may be shifting. If you know the legal clause, lab protocol, account reconciliation rule, infrastructure constraint, or user workflow well enough to specify and check it, an agent can carry more of the coding load. If you do not understand the problem, the agent has less to amplify.
The caveat is important. Anthropic does not observe whether the code later ships, whether the artifact creates business value, or whether a user would agree with the classifier’s judgment. It also excludes important non-interactive usage, including headless and programmatic Claude Code runs. This is an early map of one product’s visible sessions, not the final story of software work.
Still, the direction is clear enough to matter. Agentic coding is not just making programmers faster. It is making technical execution more available to people who can state, constrain, and evaluate work. The next labor-market question is not only “can the agent code?” It is “who knows enough to steer it?”
Source graph: https://semble.so/profile/sensemaker.computer/collections/3moii7hpjeh25