Reading Anthropic's Report on How People and Claude Code Divide the Work Across 400,000 Sessions
I read "Agentic coding and persistent returns to expertise," published by Anthropic's economics research team in June 2026. It analyzes about 400,000 Claude Code sessions from October 2025 to April 2026, using a privacy-preserving method, to see who decides what and which sessions succeed. Here are four insights from it.
Before we start — key terms
Harness — the frame around a model that lets it read files, edit code and run commands. Claude Code itself is a harness for the Claude model. It comes up in Insight 4.
Judged success / verified success — the report measures success at two levels. Judged success is when a classifier reads the transcript and decides the user did what they set out to do. Verified success additionally requires at least one hard signal, such as a commit, a pull request, passing tests or the user's explicit confirmation; it is the strictest measure. Success rates in Insight 3 are verified success unless noted.
Expertise rating — a five-point score, from novice to expert, that Claude assigns from the transcript for how skilled the user is at that task. It looks at how precisely the user frames directions, what they ask Claude to verify, and who tends to correct whom. It is rated per task, not by job title, so a senior engineer asking their first Rust question counts as a Rust novice. It comes up in Insights 2 and 3.
Insight 1: People make 70% of the planning decisions, Claude makes 80% of the execution decisions
On average, people make about 70% of the planning decisions but only 20% of the execution decisions (Figure 2). In practice, there is a clear division of labor in agentic coding––people decide what to build, and the agent decides how to build it.
Planning decisions are what to do, which approach to take and what counts as done. Execution decisions are which files to change, what code to write and which commands to run. When the user keeps control of execution (over 80% of execution decisions), Claude takes about eight actions per turn; when Claude takes control of planning (over 80% of planning decisions), it takes the most, about 16.
Claude already makes about 80% of the execution decisions. Looking at that, I think people will end up being the ones who handle the planning.
Insight 2: The more expertise a user brings, the more Claude does per instruction
In typical novice sessions, each prompt sets off about five Claude actions and roughly 600 words of output, while expert sessions set off action chains more than twice as long (12 actions) carrying five times the output (3,200 words) (Figure 3). This gap between novice and expert sessions appears within every kind of work and every band of task value.
These are split by the expertise rating described above. To show how each level gives instructions, the report includes a table of examples drawn from SWE-chat, a public dataset of coding agent sessions. Sessions rated novice give generic instructions with no domain knowledge showing, while expert sessions carry deep knowledge of the codebase and the technical environment.
If one instruction from an expert moves the work more than twice as far, I want to see for myself how the people who are really good at this actually use Claude Code.
Insight 3: What decides success is how well you know the domain, more than how well you code
In sessions that produce code, every major occupation succeeds at rates within a few points of those in software-related occupations. It appears that coding agents are making a coding background less relevant to successful programming. (...) At the same time, successful sessions are more likely to exhibit domain expertise.
Comparing sessions with the same work mode, task value, month, subject and occupation group, novice sessions reach verified success 15% of the time and intermediate-or-higher sessions 28–33%. Most of the gap comes between novice and intermediate. When a session hits trouble, 19% of novice sessions are abandoned, against 5–7% for everyone else. Software occupations do still lead slightly, at about 30% verified success against about 26% for other occupations, and that gap held steady over the seven months.
In my earlier post on the ATT&CK report, I felt that it still takes someone with real expertise to turn AI's output into a finished product, and this data points the same way, so I have to agree. That's why I think an IT degree alone, knowing how to code without knowledge of another field, will make it hard to survive.
Insight 4: Less fixing, more operating, analysis and writing
The composition of the work done with Claude Code changed substantially between October 2025 and April 2026. The clearest change is that the share of sessions spent fixing broken code fell from 33% to 19% (Figure 4). In its place, we saw a greater share of the work that surrounds code.
Operating software (deploying, configuring, running pipelines, monitoring) grew from 14% to 21% of sessions, and writing and data analysis roughly doubled, from about 10% to 20%. The estimated value of a session, priced against freelance job postings, rose 27% over the same period: 43% for building, 34% for operating and 32% for fixing. The authors say these estimates are coarse and meant for comparing tasks with one another. In a survey of social scientists from the same period, only 20% used coding agents regularly (post on that survey); this report shows how use is widening among the people who already do.
I think the shift away from fixing and toward handing over wider work comes from the models getting better, and from the harness, Claude Code itself, getting better too.
Among the signals to watch, the report notes that if the returns to expertise begin to decrease, that would suggest models are starting to supply the essential judgment that users currently bring.
