Anthropic Study on Agentic Coding: Why Returns to Expertise Persist in the AI Era

Forum topic · QianXun · 2026-06-17

Summary

Anthropic's Economic Research Center published 'Agentic Coding and Persistent Returns to Expertise' (June 2026), analyzing roughly 400,000 Claude Code sessions from 235,000 users between October 2025 and April 2026. The study finds that human expertise has not been devalued by AI coding tools—it remains a critical lever. Users make about 70% of planning decisions (what to build) while Claude handles about 80% of execution decisions (how to build it). Verified success rates rise with expertise: 15% for novices versus 33% for experts, and recovery from setbacks diverges sharply—experts complete 15% of previously failed sessions versus 4% for novices. Expertise is task-specific: domain knowledge in a given task matters more than general programming credentials. Success rates across occupations are converging within 7 percentage points of software engineers, suggesting coding is becoming a universal professional skill. Over seven months, fixing code fell from 33% to 19% of tasks while data analysis and documentation doubled, indicating Claude Code is shifting from code editor to workflow executor. The authors flag sample bias, unquantified economic value, and excluded non-interactive usage as caveats.

Anthropic's Economic Research Center published a paper titled *Agentic Coding and Persistent Returns to Expertise* on June 16, 2026. Based on real-world data from roughly 400,000 Claude Code sessions by 235,000 users over seven months (October 2025 – April 2026), it addresses a core question: when AI can write code, does human expertise still matter?

The surprising answer: expertise has not depreciated—it is more important than expected.

Source: https://www.anthropic.com/research/claude-code-expertise

Methodology

Anthropic used its Clio privacy-analysis framework for de-identification, then had Claude Sonnet 4.6 read each session transcript, cross-validated with telemetry data. A decision-attribution classifier identified every "meaningful decision," splitting them into planning decisions (what to do) and execution decisions (how to do it), attributing each to either the user or Claude.

Key Data

| Dimension | Value | |---|---| | Total sessions | ~400,000 | | Users | ~235,000 | | Time window | Oct 2025 – Apr 2026 | | Typical turns per session | ~4 | | Avg. Claude operations per session | ~10 | | Avg. output length | ~2,400 words |

Human–AI Division of Labor

Users make on average 70% of planning decisions (what to build, which approach, what "done" means)
Claude makes on average 80% of execution decisions (which files to change, what code to write, which commands to run)

This boundary is stable: humans own "understanding the problem," AI owns "using the tools."

Returns to Expertise

| User level | Verified success | At least partial success | |---|---|---| | Novice | 15% | 77% | | Intermediate | 28% | 91% | | Advanced/Expert | 33% | 92% |

The key difference lies in recovery from setbacks:

Novice sessions: only 4% verified success after being blocked
Expert sessions: 15% verified success after being blocked

Novices abandon 19% of failed sessions; experts only 5–7%. Experts don't just perform better—they know how to work around errors.

Notably, expertise is task-specific: a senior engineer asking their first Rust question is still a novice in that task; an accountant who doesn't know Python but can precisely describe reconciliation rules and spot edge cases is an expert in that task.

Convergence Across Occupations

Task success rates across occupations are within 7 percentage points of software engineers; management roles are slightly higher (possibly due to more precise expression). Software development is becoming a universal professional skill rather than a programmer-exclusive one.

Trends Over Seven Months

Fixing code fell from 33% to 19% of sessions—use is shifting toward end-to-end tasks
Operating software rose from 14% to 21%
Data analysis and document writing roughly doubled from ~10% to 20%

Claude Code's role is shifting from "code editor" to "workflow executor."

Why It Matters

For developers: AI won't take your job, but it will widen the gap between peers. The deeper your expertise, the greater the leverage AI provides. Those who ask good questions will be worth more than those who write code.
For enterprises: Training employees to "use AI to code" is the entry-level move; the scarce, valuable skill is "posing problems AI can understand," which requires business understanding.
For AI product design: Copilot-style autocomplete has peaked. Next-generation products hinge on genuinely understanding users' domain intent.
For labor markets: Software engineers' moat lies not in writing code but in judging which code is worth writing. Coding agents are replacing implementation work while rewarding deep understanding of the problem being solved.

Caveats and Open Questions

1. Sample bias: Claude Code users skew technical and senior; conclusions may not extend to non-technical users. 2. Task value unmeasured: The study cannot tell whether generated code was actually used or created economic value. 3. Non-interactive usage excluded: A significant share of API and automation use cases were not covered, where returns to expertise may differ. 4. Durability uncertain: The authors explicitly flag whether returns to expertise decline—and whether non-engineer success rates keep rising—as ongoing monitoring signals, acknowledging the conclusions could reverse.

One-line takeaway: In the AI coding era, domain expertise hasn't been devalued—it has become the most valuable lever. Those who can ask good questions are scarcer than those who can write code.