AI-Wiki

Agentic coding and persistent returns to expertise

8 min read

Agentic coding and persistent returns to expertise

The agentic-coding installment of the Anthropic Economic Index, published by Anthropic (16 June 2026), based on a privacy-preserving analysis of ~400,000 interactive Claude Code sessions from ~235,000 people, October 2025 – April 2026. Authors: Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Ryan Heller, and Peter McCrory. It extends the program’s earlier Claude Code autonomy work from how autonomous the agent is to who decides what, who is doing the work, and whether it succeeds — and frames coding as “a leading case” for where agentic knowledge work is headed.

Key findings

  • A clear division of labor: people decide what, the agent decides how. Using a privacy-preserving decision-attribution classifier, the report splits each session’s decisions into planning (what to do, which approach, what counts as done) and execution (which files, what code, which commands). On average people make ~70% of planning decisions but only ~20% of execution decisions. “People decide what to build, and the agent decides how to build it.”
  • Action volume tracks who holds control. A typical session is ~4 turns; each user prompt sets off ~10 Claude actions on average (the tail is long — ~2% of sessions average >100 actions/prompt). When the user keeps execution control (>80% of execution decisions), Claude takes ~8 actions/turn; when Claude controls planning (>80%), it takes ~16. Claude writes ~2,400 words of output per turn.
  • Returns to expertise persist — but it’s domain expertise, not coding skill. A task-specific expertise classifier (novice→expert, five-point) keys on how precisely the user frames directions, what they ask Claude to verify, and who corrects whom. Expertise is explicitly not job title and not general ability: “A senior engineer asking their first Rust question is a beginner at Rust.” More expertise → Claude does more per instruction: novice sessions ~5 actions / ~600 words per prompt; expert sessions ~12 actions / ~3,200 words (+9% actions, +13% output per expertise level, robust to controls; p < 0.001).
  • Competence captures most of the benefit; mastery adds little. Verified success (judged success plus a hard verifiable signal — commits/PRs, passing tests, explicit user affirmation) rises from 15% (novice) to 28–33% (intermediate and up); partial success 77% → 91–92%. Most of the gain is novice→intermediate; the intermediate→expert slope flattens. Among troubled sessions, verified success rises 4% (novice) → 15% (expert), and novices abandon ~19% of troubled sessions vs 5–7% for everyone else — steering is the expertise premium.
  • Occupation matters less than expertise. Occupation is inferred from transcript signals (project context, file names, referenced artifacts, vocabulary), mapped to 23 BLS SOC major groups, and explicitly not inferred from the act of coding. In code-producing sessions, every one of the ten largest occupations lands within seven points of software engineers on success (≈34% vs 29%); the gap is small and has neither widened nor narrowed. Management occupations score slightly above software engineers (possibly transferable delegation skill, possibly a measurement artifact of managers confirming success more explicitly). “Coding agents are making a coding background less relevant to successful programming.”
  • Work composition shifted in seven months. Sessions spent fixing broken code fell 33% → 19%; operating software grew 14% → 21%; writing and data analysis roughly doubled (~10% → ~20%). Usage moved toward more end-to-end agentic use (deploying/running code, analyzing data, non-code documents). ~56% of sessions write/fix/test/orchestrate code; ~17% operate software; ~14% plan or explore; ~13% produce analysis or prose. ~⅕ of sessions touch no codebase at all.
  • Task value rose ~25–27%. Each session’s value is approximated against a public dataset of freelance-marketplace postings (relative, not literal dollars). Average session value rose ~27% Oct→Apr; building, operating, and fixing tasks each grew ~⅓ or more (≈43%, 34%, 32%).

The nine modes of work

Each session is classified into the single mode that best describes its goal: building new code, fixing broken code, testing, orchestrating agents/pipelines, operating software (deploy/configure/run/monitor), understanding an existing system, planning a change, analyzing data, and communicating (presentations/prose). The classifier (Claude Sonnet 4.6) agrees with automatic telemetry on >90% of code-change labels.

How success is measured (worth importing)

The report cannot observe real-world outcomes, so it relies on transcript-based proxies: judged success (a classifier reads the full transcript: succeeded / partially / failed / no clear goal), then verified success = judged success and at least one hard signal from a success-signal classifier (git commits/PRs matching the work, passing test suites, explicit user affirmation; scored 1–5) with a parallel failure-signal classifier (errors, failed tests, retries, user pushback). Sessions with “no clear goal” (~7.7%) are excluded from the success analysis. This is a reusable scaffolding for evaluating agent sessions where ground-truth outcomes are unobservable.

Looking ahead (the authors’ framing)

Anthropic frames the report as an early read on labor-market transitions: agentic coding amplifies domain knowledge while substituting for implementation-heavy work. Two leading indicators they will track: (1) if returns to expertise begin to fall, that would signal models are starting to supply the judgment users currently bring (gains broadening beyond domain experts); (2) if non-software occupations’ success keeps rising, software production may be becoming ordinary work in every field. “Coding is a leading case — what happens in software is likely a preview of what may come as agentic tools take on other forms of knowledge work.”

What was actually ingested

The full 18-page PDF. (Originally hand-acquired under the non-descriptive content-management name CCEconReport-G.pdf; re-acquired 2026-06-18 via the zotero-acquire channel — Zotero key N6KSNEQM — which supplied the descriptive filename now recorded in raw: and consolidated the duplicate. Zotero mis-typed the item journalArticle; it is a research report, routing corrected to raw/reports/.) Figures 1–6 were read from their captions and surrounding text, not reproduced. The Appendix (classifier full text, validation results, regression details, task-estimator construction) is published separately (“Available here”) and was not ingested; classifier-validation claims here rest on the body’s summary. All classifiers use Claude Sonnet 4.6 unless noted. The report excludes non-interactive / headless (claude -p) and third-party-IDE/SDK usage — “a substantial share of activity” the authors flag as future work.

Why this matters to the wiki

This report sharpens several live threads:

  • It is the division-of-labor primary for automation-vs-augmentation at the task-within-session grain — the planning/execution split operationalises “who is doing the work” more finely than the AEI’s Directive/Feedback-Loop vs Iteration/Learning/Validation taxonomy.
  • It is the strongest evidence yet for the “domain expertise, not coding proficiency” claim that durable-skills has been assembling — and it empirically echoes Argenti’s mindset-over-skillset thesis.
  • It complicates the ai-deskilling story: the AEI’s Q4 report found first-order deskilling (the remaining human task mix has lower education content as AI covers high-education tasks), whereas this report finds that domain understanding becomes more pivotal, not less. The two are not strictly contradictory (deskilling is about task composition within a job; returns-to-expertise is about who succeeds with the tool) but they pull in opposite rhetorical directions — captured as a contradicts edge with a clarifying via.
  • It corroborates the occupation-blurring direction of the observed-exposure report (same authors) and the augmentation-dominant reshaping of BCG’s reshaping report from the usage side.

Dynamic-capabilities reading

  • digital-sensing/digital-scouting — Anthropic scouts emerging labor-market and knowledge-work signals directly from agent usage telemetry (work modes, decision attribution, success rates), a usage-data sensing instrument for where agentic work is heading.
  • digital-transforming/redesigning-internal-structures — the planning/execution division of labor is, in effect, a template for how knowledge work is being restructured around human-agent collaboration: humans concentrate on framing and verification, agents on implementation.
  • strategic-renewal/business-model — the “software production may become ordinary work in every field” framing implies business-model-level renewal of who performs technical work and how technical capability is sourced.
  • contextual/external-triggers — the agentic-coding adoption wave (GitHub coding-agent activity more than doubled since late 2025; Claude Code users averaging ~20 hours/week) is the external trigger reshaping the knowledge-work context this report reads.

Linked entities and concepts

Relationships

Relationships

Part of

  • Anthropic Economic Index — the agentic-coding installment of the AEI; extends the prior Claude Code autonomy work and the AEI usage reports to who-decides-what and success-by-expertise

Supports

Contradicts

  • Anthropic Economic Index: New building blocks for understanding AI use — tension within the AEI program — the Q4 report's first-order deskilling (remaining task mix has lower education content as AI covers high-education tasks) sits against this report's finding that returns to domain expertise persist and even sharpen; different measures, partially opposing readings

Published by

  • Anthropic — Anthropic research note; part of the Anthropic Economic Index program; based on ~400,000 Claude Code sessions

Graph View

Backlinks