On June 16, 2026, Anthropic published "Agentic coding and persistent returns to expertise," a study of roughly 400,000 interactive Claude Code sessions from about 235,000 people between October 2025 and April 2026. Its central finding is that success with a coding agent depends far more on how well a person understands the problem than on whether they are a professional programmer. People made most of the planning decisions while Claude made most of the execution decisions, and users with more domain expertise got more work done per instruction and succeeded more often. For software teams, the study is an early, data-heavy look at how the developer's job is shifting from writing code toward specifying and verifying it.
Key Facts#
- Published: June 16, 2026, by Anthropic researchers Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Shaoyi Zhang, Ryan Heller and Peter McCrory.
- Data: about 400,000 interactive sessions from about 235,000 people, analyzed with Anthropic's privacy-preserving tooling. Headless, SDK and third-party IDE usage was excluded.
- Division of labor: users made about 70% of planning decisions, while Claude made about 80% of execution decisions.
- Work per prompt: a novice prompt triggered about five Claude actions and 600 words of output, and an expert prompt about 12 actions and 3,200 words.
- Success rates: novice-rated sessions reached verified success 15% of the time, compared with 28% to 33% for sessions rated intermediate or higher.
- Occupation gap: in sessions that produced code, verified success was 34% for software-related occupations and 29% for others.
- Shifting work: the share of sessions fixing broken code fell from 33% to 19%, while operating software rose from 14% to 21%.
- Task value: the estimated value of the average session rose 27% over the period, based on a coarse comparison with freelance job postings.
What Happened#
The researchers classified each session into one of nine work modes. Writing new code and fixing broken code each accounted for about a quarter of sessions, operating software (deploying, configuring and monitoring) for 17%, planning or exploring for 14%, and data analysis or prose for 13%. A classifier then listed the meaningful decisions in each session and attributed each one either to the user or to Claude. The split was clear: people decide what to build, and the agent decides how to build it.
Expertise was rated per task, not per person. A senior engineer asking a first Rust question counts as a novice at that task, while an accountant who precisely specifies reconciliation rules and catches an edge case counts as an expert. The classifier looked at how precisely users framed instructions, what they asked Claude to verify, and who corrected whom.
Success was measured two ways. "Judged success" came from a classifier reading the transcript. The stricter "verified success" also required hard evidence, such as commits, pull requests, passing tests or explicit confirmation from the user. With controls for work mode, task value, month, subject and occupation group, expertise predicted success consistently. Most of the gain came from moving from novice to intermediate, and the step from intermediate to expert added comparatively little. When sessions ran into trouble, novices abandoned them without writing any code 19% of the time, against 5% to 7% for everyone else.
The occupation results may be the most striking. Among the ten largest occupation groups, every one landed within seven percentage points of software engineers on verified success in code-producing sessions. Management occupations came out slightly ahead, which the authors say may partly reflect managers explicitly confirming when they get what they asked for.
Background#
The study builds on Anthropic's earlier look inward. In December 2025, Anthropic reported on a survey of 132 of its own engineers and researchers, who said they used Claude in 60% of their work, with a self-reported productivity boost of 50%. Most said they could fully delegate only a small share of their work, typically favoring tasks that are easy to verify. The new report moves from self-reports to behavioral data from outside users.
Adoption context matters too. The authors cite research showing that the share of GitHub projects with coding-agent activity more than doubled after late 2025, and say Claude Code users now run the tool for an average of 20 hours per week. For a different methodology, a controlled experiment rather than usage analysis, see our coverage of METR's study of experienced open-source developers.
Why It Matters for Developers#
The findings describe where human effort pays off when an agent writes most of the code.
- Invest in the "what," not only the "how." Planning decisions stayed with people, and precise framing correlated with success. Clear acceptance criteria, constraints and examples in the prompt are the highest-leverage inputs you control, as our guide to AI-assisted .NET development explains.
- Give the agent verifiable targets. Verified success depended on hard signals such as passing tests and commits. In a .NET codebase, a fast
dotnet testloop and good integration tests turn an agent's claims into evidence it can check itself. - Domain knowledge compounds. Experts got more than twice as many actions and five times the output per prompt. Teams should pair agents with people who understand the business rules, and engineers should keep deepening domain knowledge rather than only tool skills.
- Expect more code from non-developers. If analysts, designers, scientists and managers succeed at close to engineers' rates, more code will reach your systems from outside engineering. Code review, CI gates and least-privilege credentials matter more, not less.
- Plan for agents that operate systems. Operating software grew fastest, so agents increasingly deploy and run code. Scope their permissions accordingly, as our guide to responsible AI and LLM security describes.
Keep the caveats in view. This is a vendor study of its own product, success is inferred by classifiers built on Claude Sonnet 4.6, real-world outcomes such as whether code shipped are not observed, and task values are rough estimates.
What's Next#
Anthropic says it will keep updating these measures as models and users change, and that measuring non-interactive usage, such as agents embedded in pipelines, is a priority. The authors name two signals to watch. If the returns to expertise shrink, models may be starting to supply the judgment that users now provide. If users outside software occupations keep gaining success in coding sessions, software production may become part of ordinary work in every field. Either shift would change who benefits most from coding agents, which makes both worth tracking for anyone planning a career or a team structure.