Domain expertise in agentic work

Domain expertise in agentic work is the worker capability to frame the right problem, set acceptance criteria, recognize edge cases, and steer an AI agent toward useful execution even when the worker is not personally performing every technical step. The construct is distinct from formal coding skill, AI literacy, or generic receptivity: it is task-specific command of what good work means in context.

2026-hitzig-agentic-coding-expertise gives this topic its first direct evidence. In roughly 400,000 Claude Code sessions, the more task-specific expertise users displayed, the more work Claude did per prompt and the more likely sessions were to succeed. Expert sessions averaged about 12 Claude actions and 3,200 words of output per prompt, compared with about 5 actions and 600 words for novice sessions; novice sessions reached verified success 15% of the time, while intermediate-or-higher sessions reached 28–33% verified success.

The key I-O interpretation is that agentic AI can reduce the need to personally execute some technical steps without removing returns to knowing the work. In Hitzig et al.’s data, software-related occupations were only modestly ahead of other professions in code-producing sessions, while task-specific expertise remained predictive of success. That suggests a redesign pathway where occupational boundaries around implementation weaken, but construct-relevant domain judgment becomes more important as the scarce human contribution.

Connections

• Connects to agentic-delegation because delegation quality depends on whether the human can specify goals, constraints, and verification criteria before handing execution to the agent.
• Connects to human-ai-task-allocation because task-specific expertise is a human-state variable that should affect whether the human plans, executes, reviews, or escalates a workflow turn.
• Connects to work-redesign because agentic workflows may shift work from direct implementation toward problem framing, review, acceptance criteria, and recovery from agent errors.
• Connects to ai-mediated-learning because the same pattern that rewards expertise may also change where novices get the practice needed to become competent agent directors.
• Connects to construct-validity because “expertise,” “coding skill,” “AI literacy,” and “occupation” are separable constructs and should not be collapsed when measuring agentic-work capability.

Contradictions & open questions

• 2026-hitzig-agentic-coding-expertise suggests most gains come from moving from novice to intermediate rather than from intermediate to expert, which complicates a simple “deep expertise wins” story; working competence may capture much of the agentic-use benefit.
• The report measures apparent task expertise from transcripts, so the wiki still needs independent validation of expertise ratings and evidence that transcript-coded success corresponds to durable real-world outcomes.
• If agentic tools let non-coders produce technical artifacts, organizations will need to decide which expertise should be selected, developed, certified, or paired with review rather than assuming coding occupation is the right boundary for technical work.