Why Coding Agents Do Not Eliminate Expertise — with Evgeny Sergeev

Alexander Polomodov and Evgeny Sergeev examine Anthropic’s study of 398,198 Claude Code sessions from about 234,751 users between October 2025 and April 2026. It is a rare view into product telemetry, but it represents one tool’s audience and only the behavior captured in its logs.

Expertise is local to a task rather than guaranteed by title or tenure. A senior engineer can be a novice in Rust, while a domain specialist may automate a familiar workflow successfully. Knowledge of constraints, correctness criteria, and the intended outcome matters more than a broad user label.

Experts provide more specific context, delegate longer chains, request verification more often, and recover better after failure. Their sessions average about 12 agent actions and 3,200 output words versus 5 actions and 600 words for novices; verified success rises from roughly 15 percent to 28–33 percent.

Task value and success are proxies, and telemetry cannot reveal production impact, incidents, rework, or causality. Agent output therefore needs independent tests, acceptance criteria, decision logs, and evidence. Coding agents convert expertise into throughput only when an accountable person understands the result and accepts its risk.