Event date: June 16, 2026 Source: Anthropic Research — original article: https://www.anthropic.com/research/claude-code-expertise
Anthropic's Economic Research Center has published a paper, *Agentic coding and persistent returns to expertise*, based on real-world data from approximately 400,000 Claude Code sessions, 235,000 users, and seven months of usage (October 2025 to April 2026). It tackles a central question: when AI can write code, does human expertise still matter?
The answer is surprising: expertise hasn't depreciated—it matters more than expected.
Methodology
Anthropic used its Clio privacy-analysis framework for anonymization, then had Claude Sonnet 4.6 read each session transcript and cross-validate against telemetry. A decision-attribution classifier identified every "meaningful decision," splitting them into planning decisions (deciding what to do) and execution decisions (deciding how to do it), and attributing each to the user or to Claude.
This enabled both a whole-session view and precise per-instruction attribution—far more granular than most prior research on AI usage patterns.
Key data
| Dimension | Value | |------|------| | Total sessions | ~400,000 | | Users | ~235,000 | | Time window | Oct 2025 – Apr 2026 | | Typical session turns | ~4 | | Average Claude operations per session | ~10 | | Average output length | ~2,400 words |
Humans set direction, AI walks the path
- Users made on average 70% of planning decisions (what to do, which approach, what counts as done)
- Claude made on average 80% of execution decisions (which files to change, what code to write, which language, which commands to run)
- Novice sessions: only 4% achieved verified success after encountering difficulty
- Expert sessions: 15% did
- Code fixing fell from 33% to 19% of usage—use cases shifting toward end-to-end work
- Operating software rose from 14% to 21%
- Data analysis and documentation writing doubled from ~10% to 20%
- For developers: AI won't take your job, but it will widen the gap between you and your peers. Deeper expertise means greater leverage—those who can ask good questions become more valuable than those who can merely write code.
- For enterprises: Training employees to "use AI to write code" is the entry-level move; the real value lies in training them to "frame questions AI can understand," which requires business understanding and is far scarcer.
- For AI product design: Copilot-style "code completion" has peaked. The next battleground is "truly understanding the user's domain intent"—embedding domain knowledge graphs and workflow context into models.
- For the labor market: Software engineers' moat isn't writing code but judging which code is worth writing. Coding agents are replacing implementation work while rewarding deep understanding of the problem being solved.
This division of labor was remarkably stable: humans own "understanding the problem," AI owns "using the tools."
The most counterintuitive finding on returns to expertise
| User level | Verified success | At least partially successful | |---------|---------|------------| | Novice | 15% | 77% | | Intermediate | 28% | 91% | | Advanced/expert | 33% | 92% |
The key difference lies not in peak success rates but in recovery from setbacks:
Experts don't just do better—they know how to route around errors. 19% of failed novice sessions ended in abandonment, versus only 5-7% for experts.
Notably, expertise is task-specific—a senior engineer asking about Rust for the first time is still a novice; an accountant who doesn't know Python but can precisely describe reconciliation rules and spot edge cases is an expert at that task.
Occupational gaps are narrowing
Task success rates across occupations were within 7 percentage points of software engineers; management roles were slightly higher (possibly due to more precise expression). This supports the view that software development is becoming a general professional skill rather than a programmer-only one.
Trends over seven months
In other words, Claude Code is evolving from "code editor" to "workflow executor."
Why it matters
Risks and open questions
1. Sample bias: Claude Code users skew technical and senior; conclusions may not generalize to non-technical users writing code with AI. 2. Task value is hard to quantify: The study cannot determine whether generated code was actually used or produced economic value. 3. Non-interactive usage excluded: A significant share of API calls and automation scenarios weren't covered; the "returns to expertise" curve may differ there. 4. Durability in question: The authors explicitly flag whether returns to expertise decline and whether non-software occupational success rates keep rising as signals to monitor—implying the conclusions could reverse in the future.
One-line takeaway: In the age of AI coding, domain expertise hasn't lost value—it has become the most valuable lever. People who can ask good questions are scarcer than people who can write code.