Claude Code usage data shows domain expertise, not coding skill, drives success

This report from Anthropic analyzes ~400,000 Claude Code sessions to understand how agentic coding is actually used in practice. The core tension it exposes is straightforward: as coding agents become more capable, does deep coding skill still matter, or does domain expertise become the bottleneck? The data suggests a clear answer. People still make roughly 70% of planning decisions (what to build), while Claude handles about 80% of execution decisions (how to build it). The gap between software engineers and other occupations in successful coding sessions is small—within seven percentage points—and every major occupation succeeds at nearly the same rate. The real differentiator is not job title but task-specific domain expertise: expert-rated sessions achieve verified success more than twice as often as novice-rated sessions, and Claude produces five times more output per prompt for experts than for novices.

The concrete methodology is worth noting. The researchers built classifiers to infer occupation from session transcripts, classify nine distinct work modes, rate user expertise on a five-point scale, and judge success using both a classifier and hard verification signals like git commits and passing tests. The results show clear trends over the seven-month study period. The share of sessions spent fixing broken code fell from 33% to 19%, while operating software, analyzing data, and writing documents all grew. The estimated economic value of the average session rose by 27%. Sessions rated intermediate or above reach verified success 28-33% of the time, compared to 15% for novices. Crucially, most of the gain comes from moving from novice to intermediate, not from intermediate to expert—competence in a domain captures most of the benefit.

The practical takeaway for builders is this: coding agents are not substituting for domain expertise—they are amplifying it. A manager, lawyer, or accountant who understands their problem deeply can now direct Claude to produce working software, even without formal coding training. The tool rewards command of the problem, not command of the syntax. Novices give up far more often when sessions hit trouble (19% abandonment rate versus 5-7% for everyone else), suggesting that the ability to steer the agent through errors is a learned skill, not an innate one. If these patterns hold, the labor market may shift to reward domain understanding over pure implementation ability. The report is candid about its limits—it cannot measure real-world outcomes or non-interactive usage—but it provides one of the most detailed empirical portraits yet of how knowledge workers actually collaborate with a coding agent.