- Author
- Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Shaoyi Zhang, Ryan Heller and Peter McCrory (Anthropic)
- Published
- 16 June 2026
- Kind
- Company research report based on usage data
- Original
- www.anthropic.com/research/claude-code-expertise
- Subject
- Review, taste and expertise
- Themes
- Skill under automation
What it is
Zoe Hitzig and five co-authors wrote this Anthropic research report. It studies interactive Claude Code sessions from October 2025 to April 2026. It sets out a framework for what work is done and whether it succeeds.
What it says
- People make about 70% of planning decisions and 20% of execution decisions.
- Expert-rated sessions set off 12 actions per prompt, against about five for novices.
- Novice-rated sessions reach verified success 15% of the time, against 28-33% for intermediate and above.
- In troubled sessions, verified success rises from 4% for novices to 15% for experts.
- Troubled novice sessions end abandoned 19% of the time, against 5-7% for others.
- In code-producing sessions, the ten largest occupations land within seven points of software engineers.
- The share of sessions spent fixing broken code fell from 33% to 19%.
What evidence it gives
Claude Sonnet 4.6 classifiers label about 400,000 sessions from about 235,000 people. Comparisons hold work mode, task value, month, subject and occupation group fixed. Over 90% of sessions labelled as code-creating showed code changes in telemetry. Success is judged from transcripts, using signals such as commits and passing tests.
Interests and limits
Anthropic wrote the report about its own product. The authors call the findings preliminary. Real-world outcomes, such as whether code was used, are not observed. All labels depend on a model reading transcripts. Headless and third-party usage is excluded. The main report was read in full, but the linked appendix was not fetched.
Connections
These links are this wiki's own reading, not claims the sources make.
- Agentic Coding is a Trap. Faye worries that overuse erodes the skill that supervising agents requires. Hitzig's report shows that skill still pays, with novices reaching verified success far less often. The report covers seven months. Tracking the same people for longer would test Faye's worry directly.