State of the Article
TLDR Domain expertise, not coding skill, decides how much a coding agent can do for you.
The takeaway: when you work with a coding agent, knowing your field matters more than knowing how to code.
Anthropic analyzed 400,000 Claude Code sessions between October 2025 and April 2026 and found that a user's domain expertise, not their programming ability, decided how much the model could do on its own. Users with deep field knowledge triggered action chains twice as long as novices (12 steps versus 5) and got five times the output per instruction.
The division of labor is telling. Users made roughly 70 percent of the planning decisions while Claude handled 80 percent of the execution, a clean example of human-in-the-loop work. Across law, accounting, design, and data analysis, success rates on coding tasks converged near professional software-engineer levels. In other words, agentic coding rewards people who know what they are building, even if they cannot write the code themselves.
The mix of work shifted over the period too. Debugging fell from 33 to 19 percent of sessions while higher-value tasks grew: deployment and data analysis doubled, software operation rose from 14 to 21 percent, and the estimated economic value of an average session climbed 27 percent.
The pattern suggests these tools amplify expertise more than they replace engineers. For how that hand-off actually works, see the Agents Guide.
Source: Anthropic