Can people without coding training direct a coding agent through technical work, and what do the people who succeed bring to it? Anthropic's Economics team addresses this in a report published on 16 June 2026, whose page carries the browser title "How Claude Code is used in practice" and is titled "Agentic coding and persistent returns to expertise" on the page, in the PDF and in the suggested citation. The report classifies about 400,000 interactive Claude Code sessions with a Claude model and reads three things from them: what kind of work each session does, who decides what, and whether it succeeds. Its headline claim is that domain expertise, not coding background, goes with success, and that most of the gain comes between novice and intermediate users. The authors call the findings "preliminary" ("Looking ahead").
Versions. The version of record is the web page with its 18-page PDF (section numbers below are the PDF's) and a 27-page appendix PDF that prints the classifier prompts, validation and regression tables. Two appendix files are linked from the page; their text differs by one character. The suggested citation lists six authors (Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Shaoyi Zhang, Ryan Heller and Peter McCrory); the PDF's byline omits Shaoyi Zhang. The web page has one sentence the PDF lacks, on partial success by occupation (see below).
What the report does
Sample
The appendix ("Sample") describes 398,198 sessions sampled uniformly at random from Claude Code traffic between October 2025 and April 2026, from 234,751 users; 73,066 users appear more than once, 6,382 more than five times and 1,413 more than ten times. Internal Anthropic use is excluded. The sample covers the command-line interface, the desktop app and claude.ai. It excludes third-party IDE integrations, software development kits and "headless" single-prompt calls (claude -p), which footnote 4 describes as largely programmatic; the authors call this excluded usage "a substantial share of activity" ("Looking ahead").
Classifiers and privacy
Every label comes from a Claude model reading the transcript inside Anthropic's privacy-preserving analysis infrastructure; researchers do not read transcripts, and every reported cell must contain a minimum number of distinct users (appendix, "Sample"; footnote 7). Almost every classifier is a single call to Claude Sonnet 4.6 at temperature 0.2 choosing from a fixed list. Each turn is truncated in the middle at 5,000 characters and the whole transcript at 25,000 characters, a limit that long sessions will exceed. Telemetry recorded for every session (turns, model calls, tool calls, output tokens, lines of code added and removed, account age) is used alongside the labels.
The report's main constructs are defined as follows:
- Work mode: one of nine labels for the session's dominant activity (Plan, Build, Fix, Understand, Test, Operate, Analyze, Orchestrate, Communicate), chosen "by DOMINANT activity" rather than by the user's opening request (appendix, "Work Mode").
- Decision attribution: a classifier lists the meaningful decisions in a session, sorts them into planning (what to do) and execution (how to do it), and attributes each to the user or to Claude (section 2.2).
- Expertise: a five-point rating of the user's apparent expertise at the task at hand, from three co-equal signals: how specifically the user frames the problem, what kind of verification they ask for, and who corrects whom on domain matters (section 2.3; appendix, "User expertise").
- Occupation: one of the 23 major groups of the US Bureau of Labor Statistics' Standard Occupational Classification, inferred from side signals such as CLAUDE.md files, file names and vocabulary, with instructions not to treat coding itself as evidence of a software profession; about 30% of sessions are left "Unclear" (section 3.1; appendix).
- Task value: Claude Haiku 4.5 rewrites the session as a freelance job posting, and Claude Opus 4.7 prices it against 200 real fixed-price postings, 50 from each of four price tiers (appendix, "Estimated task value").
- Success: a judged-outcome classifier (succeeded, partially succeeded, failed, no clear goal) plus two 1-to-5 scales for the strength of success and failure evidence. Verified success requires a judged success and a success-signal score of 4 or 5, meaning at least one hard signal such as a matching commit, a passing test, a command completing with the requested output, or explicit user affirmation. A session hits trouble when the failure signal is 3 or higher, and a troubled session is abandoned when it is judged failed and telemetry shows no lines of code added (appendix, "Derived measures and definitions"). Sessions judged to have no clear goal, about 7.7%, are excluded from the success analysis (section 4).
The classifier prompts
The appendix prints every prompt in full. Some instructions shape what the labels can mean. The work-mode prompt has no "Unclear" option and tells the model to pick the best fit "even when evidence is thin". The judged-success prompt folds hard and soft signals into one label and tells the model, for thin transcripts, to "default to Succeeded if the trajectory is clean and the agent is on-task". The report's stricter measure, verified success, needs a hard signal as well; this note reads that as a guard against the lenient default, though the report does not say so. The occupation prompt says that writing code "is NOT evidence they are a software developer by profession" and should yield "Unclear" without other signals. The expertise prompt asks the model to rate the user "on the task AT HAND, not their career". The report does not test alternative wordings of these prompts.
Validation
The appendix ("Classifier validation and cross-check") reports three kinds of check:
- Against telemetry. More than 90% of sessions labeled as creating or modifying code show code changes in telemetry (section 2.1). The share judged successful rises from 13% at success-signal 1 to 90% at 5, and falls from 85% at failure-signal 1 to 3% at 5 (Figure A1). Sessions rated more Claude-led on planning show more model calls and tool calls per prompt (Figure A2).
- Against a stronger model. The classifiers were rerun on 198 sessions from SWE-chat, a public dataset of coding-agent sessions from public GitHub repositories. A stronger model, named in the appendix as Mythos Preview, labeled the same sessions as a reference, "treating the reference model labels as reliable and consistent proxies for ground-truth". Agreement was 78 to 98% for categorical classifiers; for ordinal scales it was 78 to 99% counting adjacent points and 53 to 68% for exact matches. A blind judge, again Mythos Preview, rated most disagreements as genuinely ambiguous; "concerning" disagreements were 0 to 3% of sessions per classifier (at most 5 of 198). One author blindly re-adjudicated the 15 concerning cases and agreed with the judge in 11.
- Against outcomes at Anthropic. On internal sessions, where commits can be traced to the main branch, the expertise-success gradient replicates within engineers, with judged success rising roughly 3 to 5 percentage points per expertise level under author fixed effects, and higher-rated sessions more often producing code that lands. No table is printed for this check.
The authors argue that human labels may not be a reliable gold standard for sessions this long and complex, and that relative comparisons are robust to classifier error as long as errors are similar across groups, a condition "which we did not extensively verify here". The SWE-chat sessions are public open-source sessions, not a sample of the study's own traffic.
What it finds
The work and the division of labor
In the full period, 26% of sessions fix something broken, 25% build something new, 17% operate software, 10% write documents or presentations, 7% each understand a system or plan a change, 3% each orchestrate agents or analyze data, and 2% test code (Figure 1). A typical session has about four user turns; each prompt sets off about 10 actions by Claude on average and about 2,400 words of output (section 2.2). The tail is long: about 2% of sessions average more than 100 actions per prompt (footnote 6).
Users make about 70% of planning decisions and Claude about 80% of execution decisions in the typical session (Figure 2), which the report summarizes as "People decide what to build, and the agent decides how to build it" (section 1). When users keep more than 80% of execution decisions, Claude takes about eight actions per turn; when Claude makes more than 80% of planning decisions, about 16 (section 2.2). Appendix Figure A2 plots actions per prompt by decision share on a different scale (roughly 17 to 44 across bands, read from the figure, which prints no values); the report does not reconcile the two scales, which may reflect different averages.
Change over seven months
Between October 2025 and April 2026, the share of Fix sessions fell from 33% to 19%, Build also fell (Figure 4 ends it at 21%, from about 27% read from the figure; the text does not mention it), Operate rose from 14% to 21%, and the text says writing and data analysis roughly doubled, from about 10% to 20% (section 3.2; Figure 4). Figure 4 labels only the Communicate line, which rises from about 10% to 20%; Analyze is not drawn. By the task-value measure, the average session's estimated value rose 27% (section 3.2; Figure 4 prints "all +27%"), with Build up about 43%, Operate 34% and Fix 32%; Figure 4 prints Communicate at +2%. The key findings round the average rise to "about 25%".
The task-value estimates are coarse by the appendix's own account. On a holdout of 999 postings across four price tiers, the pricer scores 0.38, mostly from ordering across tiers (the appendix's text calls this a Pearson correlation of log prices, "log R", while its table headers say log R² and level R², and print identical values in both columns for every tier); within-tier discrimination is "essentially zero above $5,000", small jobs are over-priced about 2.4 times and the largest under-priced about 7 times ("Validation on a stratified holdout"). The authors therefore use the values to compare sessions and months, not as dollar amounts (footnote 8). The value series also changes with the mix of models and users over the period; the report does not separate those sources.
Expertise and what Claude does per prompt
In novice-rated sessions each prompt sets off about five actions and 600 words; in expert-rated sessions about 12 actions and 3,200 words (Figure 3). In a regression controlling for work mode, task value, month, occupation and model family, each expertise level is associated with 9% more actions and 13% more output per prompt (Figure 3 caption, which gives p < 0.001 for the unadjusted trends and calls the regression estimates significant without a level). These are associations: the report does not show whether expertise causes longer action chains or whether both follow from the kind of request.
Expertise and success
Verified success rises with rated expertise. The text gives 15% for novice-rated sessions and 28 to 33% for intermediate and above; at least partial success is 77% against 91 to 92% (section 4.1; Figure 5). The appendix's Table A2 gives the underlying adjusted rates, under fixed effects for work mode by value band, month, task occupation and user occupation: 14.5% (novice), 20.9% (beginner), 28.3% (intermediate), 29.5% (advanced) and 32.9% (expert). Unadjusted, the expert rate is 34.7%. In the linear version (Table A1), each expertise step adds 3.8 percentage points to the probability of verified success in the preferred specification (column 5), against 4.3 points with no controls, with standard errors of 0.08 to 0.09 points clustered by user.
Two features of these tables bear on how far the gradient goes. Novice-rated sessions are few: 5,911 of the 359,586 in Table A2, about 1.6% (this note's arithmetic), against 146,879 rated advanced. (Table A1's caption gives 375,687 sessions, but its columns report 359,586 or fewer; the appendix does not explain the difference.) The regression's R² is small (0.009 without controls, 0.088 in column 5), so rated expertise explains little of the session-level variation in success even while its coefficient is precisely estimated.
Among sessions that hit trouble, verified success rises from 4% (novice) to 15% (expert), at least partial success is 60% for novices against 80 to 81% for intermediate and above, and 19% of troubled novice sessions end abandoned against 5 to 7% for others (section 4.1; Figure 5). Footnote 9 adds a selection caveat: experts hit trouble less often, and the estimated value of a troubled session roughly doubles from the bottom of the expertise scale to the top, so experts' troubled sessions are likely harder problems.
Occupation
Users in software-related occupations reach verified success in about 30% of sessions and other identified occupations in about 26%; in sessions that add or change code, 34% against 29% (section 4.2). The gap "has neither widened nor narrowed over seven months". In code-producing sessions every one of the ten largest occupation groups is within seven points of software users (Figure 6); management is highest on verified success (37% in Figure 6 against 34% for software). The caption's "within seven percentage points" holds for verified success; on judged success, which the figure also prints, software users are at 60% and sales at 51%, a nine-point gap. The authors offer two readings: management skills may transfer to directing an agent, or managers may be more likely to confirm explicitly that they got what they asked for, which the verified-success measure counts (section 4.2; footnote 10).
The web page adds a sentence absent from the PDF: that at least partial success in code-producing sessions is 89% for software occupations and 88% for others. Figure 6, on both versions, prints at least partial success of 94% for software users and 92 to 95% for the other nine groups shown. The page does not explain the difference; the figure covers only the ten largest groups.
The occupation analysis rests on the inferred profession of the user, which is missing for about 30% of sessions (appendix). Management, sales and legal occupations were the fastest-growing non-software groups (section 3.1), but the report does not print the occupation shares.
The report's reading
In "Looking ahead" the authors write that coding agents appear to be "making a coding background less relevant to successful programming", that the ability to steer Claude "comes more from command of a domain than from the ability to write code", and that a working grasp of a domain "captures most of the benefit". They propose tracking two shifts: a falling return to expertise, which would suggest models are supplying the judgment users now bring, and a rising share of successful coding sessions from users outside software occupations.
Limitations
As stated in the report ("Looking ahead" and the appendix):
- real-world outcomes are not observed: whether code is kept, used or valuable;
- non-interactive usage, a substantial share of activity, is excluded;
- every label depends on a model's reading of a truncated transcript, and validation at scale is hard; the authors say the results "caution against over-reading any particular label";
- task values are coarse and used only comparatively.
Further limits follow from the design, as this note reads it:
- Shared signals. Expertise, success and failure are all read from the same transcript. The expertise rubric rewards a user who catches Claude's mistakes; the failure scale counts user pushback; the success scale counts explicit affirmation. The report does not test how far these signals overlap.
- Correlational design. The controls narrow comparisons to similar sessions, but rated expertise is not assigned, and the report does not claim it causes success.
- One developer, one product. These are Anthropic's measurements of its own product; the data cannot be checked outside the company.
Research questions
- Does the expertise-success gradient hold when expertise is measured outside the transcript? Rated expertise and success share signals. A study that measured users' domain expertise independently (a survey or a test) and linked it to session outcomes would show whether the gradient survives.
- How reliable is verified success against real outcomes? Commits and affirmations are proxies. The internal check on whether commits land was described but not tabulated; publishing it, or linking sessions to later code survival in external repositories, would calibrate the measure.
- Is the partial-success figure for occupations 89% or 94%? The web text and Figure 6 disagree. A correction or a table of the occupation results by success measure would settle it.
- Does the return to expertise fall as models improve? The authors propose tracking it. Repeating Table A2 by model generation, with the same classifiers, would answer it directly.
- What does the excluded non-interactive usage look like? Headless and SDK use is described as substantial. A framework for classifying it, which the authors list as future work, would show whether the division of labor differs without a user present.
- How much does the 25,000-character truncation affect the labels for long sessions? Long sessions are where expert users set off longer action chains. Comparing labels on full and truncated transcripts for a sample would measure the effect.
- Would human annotators agree with Mythos Preview on these sessions? The validation treats a stronger model as the reference. A human-labeled subset of shorter sessions, compared with both the reference and the report's classifiers, would show how close model agreement is to human agreement.
How it connects
- Measuring AI agent autonomy in practice (February 2026) is the prior Claude Code report the authors cite; it tracked how long the agent runs and how often users auto-approve actions. The report describes its decision-attribution and per-prompt measures as complements to those (section 2.3).
- How AI is transforming work at Anthropic (December 2025) is the other prior work the report cites, on Claude Code inside Anthropic.
- Anthropic Economic Index: AI’s impact on software development (April 2025) compared Claude Code with Claude.ai using Clio's automation and augmentation categories; this report uses a different taxonomy (work modes and decision shares), so its figures are not continuous with that post's. This note makes the comparison.
- Clio is the privacy-preserving analysis tool the report links to for its method.
- How AI agents shape experts’ judgment calls (October 2026) is a randomized trial on how a Claude Code agent's lean affects experts' analytical decisions, a question about expert steering that this observational report cannot test. This note makes the connection.
- AI for scientific discovery: How researchers use Claude (November 2026) applies a similar classifier pipeline, with an autonomy measure, to research sessions on Claude and Claude Code; Massenkoff, Heller and McCrory are authors of both. This note makes the connection.
- Coding agents in the social sciences (May 2026) surveys researchers' adoption of coding agents and is the self-report counterpart for one occupation group. This note makes the connection.
References
- Hitzig, Z., Massenkoff, M., Lyubich, E., Zhang, S., Heller, R., & McCrory, P. (2026, June 16). Agentic coding and persistent returns to expertise. Anthropic. https://www.anthropic.com/research/claude-code-expertise (PDF: https://assets.claude.com/433472e34b60db1a52ebf0b8c6600f057b6908c5.pdf) record
- Anthropic. (2026, June). Appendix to "Agentic Coding and Persistent Returns to Expertise" [PDF, 27 pp.]. https://assets.claude.com/a94728142a45694292336165947f8d6e3e1a357e.pdf record