"Agentic coding and persistent returns to expertise" (Hitzig et al., 2026) provides an empirical analysis of human-AI interaction patterns using ~400,000 Claude Code sessions from October 2025 to April 2026. The research explores how agentic tools transform knowledge work, emphasizing a persistent dependency on domain expertise rather than formal coding proficiency.

Core Methodology

The study employs a framework of transcript-based classifiers to analyze session dynamics. The authors utilize:

1. Decision Attribution Classifier: Analyzes session transcripts to partition decision-making into two categories:
   • Planning ($P$): Defining "what to do," the approach, and criteria for success.
   • Execution ($E$): Determining "how to do it," including file selection, syntax, and command execution.

The model outputs user-share ratios for these categories, revealing that users typically retain ~70% of planning control while delegating ~80% of execution.

2. Expertise Classifier: Rates user competence on a 5-point scale (Novice to Expert) based on framing precision, verification requests, and corrective behavior. This is task-specific rather than role-based.
3. Success Metrics:
   • Judged Success: A model-based assessment of goal attainment.
   • Verified Success: A binary indicator requiring both judged success and hard evidence (e.g., successful git commits, passing test suites).
4. Economic Valuation: Estimated task value is derived by mapping sessions to a freelance marketplace dataset, allowing for longitudinal analysis of work complexity and economic output.

Key Findings

• The Division of Labor: There is a clear bifurcation of responsibilities. Users act as architects (planning), while Claude functions as an implementer (execution). The agent's output volume scales with the user's expertise; expert-rated sessions trigger twice as many actions and five times the word count per prompt compared to novices.
• The Expertise Premium: Domain expertise is the primary driver of success. While the performance gap between intermediate and expert users is modest, the jump from novice to intermediate is significant. Novices are more likely to abandon sessions (19% vs. 5–7% for others) when encountering technical hurdles.
• Coding as a Horizontal Skill: In code-producing sessions, the success rate of non-software professionals (e.g., in Law, Management, or Finance) is within seven percentage points of software engineers. This suggests that agentic coding is lowering the barrier to technical implementation, rendering coding background less relevant than the ability to define the problem correctly.
• Evolving Usage: Over the seven-month observation period, the utility of sessions rose by approximately 25%. A notable shift occurred in session composition: the proportion of debugging work fell by nearly half (from 33% to 19%), while data analysis, documentation, and system orchestration increased, signaling a shift toward end-to-end agentic workflows.

Implications

The study concludes that agentic coding agents do not substitute for domain expertise; rather, they amplify it. The findings suggest a labor market transition where the ability to "steer" an agent effectively—requiring a strong understanding of the underlying domain problem—is becoming more valuable than the technical implementation skills traditionally associated with software engineering. Persistent returns to expertise imply that while tools broaden access to technical work, they continue to reward those who possess a firm grasp of their specific field.