Agentic coding and persistent returns to expertise
9 hours ago
- #AI-assisted coding
- #domain expertise
- #human-AI collaboration
- Framework introduced for analyzing agentic coding via ~400k Claude Code sessions from Oct 2025-Apr 2026, focusing on task composition, human-AI collaboration, and success rates.
- Human users make ~70% of planning decisions; Claude handles ~80% of execution decisions, with domain expertise increasing Claude's work per instruction.
- Success rates for coding tasks are similar across occupations, with domain expertise (not coding proficiency) driving higher success, especially from novice to intermediate levels.
- Usage shifted from debugging (down by nearly half) to end-to-end agentic work like deployment, data analysis, and non-code documents over seven months.
- Estimated economic value of typical tasks rose ~25% on average across most work types, based on comparisons to freelance job postings.
- Expert sessions generate more Claude actions and output per prompt than novice sessions, and experts recover more often from trouble.
- Occupation classification shows software-related jobs are largest group, but non-software occupations succeed at comparable rates in code-producing sessions.
- Measurement relies on transcript classifiers for expertise, success, and occupation, with privacy-preserving methods, though real-world outcomes are unobserved.
- Findings suggest agentic coding rewards domain expertise over coding skills, potentially broadening who can do technical work while emphasizing problem-solving knowledge.
- Future research will track shifts in returns to expertise and occupational success rates to understand labor market impacts.