# What Decides Whether a Coding Session Succeeds? Anthropic's New Study Says Domain Expertise, Not Coding Background. Published: 2026-06-17 **So what decides whether a coding session succeeds?** Anthropic just put out a study. Turns out it is not your coding background.. that is becoming less relevant. It is **domain expertise**. The more you understand the problem you are solving, the more the agent does per instruction, and the more often the session ends in success. Matches my experience.. 25-odd years as an analyst ....SAS, Python...from analysis to ML models...on the coding itself these agents beat me by multiples. Where I still win is the deep domain side.. greatly satisfying catching Claude when it makes a mistake while confidently telling me it validated everything. I pulled some main findings, quoted straight from the report, into the attached carousel. Full Anthropic report here: --- More on how I build and work with these agents.. over at [tigzig.com](https://tigzig.com). It's an AI-agent first site - so you can just point your **Claude Cowork, Claude Code, or any AI agent** at it and have it explore for you. Or feel free to explore yourself. --- ## Full Deck Content (Text Format) ### Slide 1 - The three biggest takeaways **Anthropic's new study · Claude Code · the three biggest takeaways** - **Coding backgrounds are less relevant.** - **Management occupations show the highest success.** - **More domain expertise, more success.** *Based on Anthropic's analysis of about **400,000 Claude Code sessions** from **235,000 people**, October 2025 to April 2026. Findings quoted in full inside.* ### Slide 2 - Coding background is becoming less relevant > "Every major occupation succeeds at rates within a few points of those in software-related occupations." Software engineers verified at **34%**. Lawyers, analysts, scientists, marketers, sales: every one of the ten largest groups lands within **seven points**. ### Slide 3 - Management occupations come out highest > "Management occupations are highest on verified success, slightly above the software engineering occupations." Anthropic's own read: their higher success *"may reflect management skills that transfer to directing an agent."* With a caveat, that it *"may also partly reflect our measurement,"* since managers may be quicker to confirm when they got what they asked for. ### Slide 4 - More expertise, more work per instruction > "The greater domain expertise a person brings to a session, the more work Claude does per instruction." Same tool, different driver. Anthropic measured how much Claude does off a single prompt, by the user's rated expertise at the task. **More than twice the actions and five times the output from the same instruction.** Expertise here is *task-specific, not job title*. ### Slide 5 - And more often the session ends in success > "The more domain expertise a person has, the more often the session ends in success." Verified success **more than doubles, 15% to 33%**. Most of the jump comes early, moving from novice to intermediate. ### Slide 6 - When trouble hits, expertise pulls the session back When a session hits trouble, expertise is what pulls it back. > "**19% of sessions where the user appears to be a novice end abandoned**, against 5 to 7% for everyone else." ### Slide 7 - The mix is shifting from fixing to building > "The share of sessions spent fixing broken code fell from 33% to 19%." **+27%** — value of the average task, versus a freelance-market benchmark. The mix is shifting from *fixing* toward *building, running, analyzing and writing*, and the typical task is worth more. ### Slide 8 - The honest caveat from Anthropic > "We cannot measure real-world outcomes... whether code written in a session is actually used or discarded." Nor whether a session *"produces an economically valuable artifact."* Every label comes from a model reading the transcript, and *"classifiers remain challenging to validate at scale."* ### Slide 9 - Where I build & write · Source & method #### Where I build & write Live macro data, database AI, quants and MCP servers, plus **45 live apps**, **200+ build guides** and a security checklist for tool builders: [tigzig.com](https://tigzig.com). **Agent-first portal.** Point Claude Cowork, Claude Code, or any AI agent at [tigzig.com](https://tigzig.com): it finds the right tool, guide or dataset, hits the APIs to pull and analyze the data, and sets up the open-source tools and MCP servers for you. #### Source & method The report: **Agentic coding and persistent returns to expertise.** Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Ryan Heller, Peter McCrory. Anthropic, 16 June 2026. [anthropic.com/research/claude-code-expertise](https://www.anthropic.com/research/claude-code-expertise) A privacy-preserving analysis of about **400,000 interactive Claude Code sessions** from about **235,000 people**, Oct 2025 to Apr 2026, classified by a model reading each transcript and cross-checked against automatic telemetry. #### What I did here Every headline is a **direct quotation**. All charts were rebuilt from Anthropic's figures (3 to 6) in this house style; all data is Anthropic's. --- *Amar Harolikar · Decision Sciences & Applied AI · [tigzig.com](https://tigzig.com).*