Coding agents lift coding speed 73.7% but delivery only 13.7%, as review becomes the bottleneck
写码提速 74%,交付只涨 14%:当人肉审代码和 AI 审查插件双双见顶,团队该怎么破局
A Harvard working paper found that within 12 months of adopting coding agents, engineers' hourly coding output rose 73.7% while shipped tickets grew just 13.7%, making code review the delivery bottleneck. Drawing on the Claude Code team's methods, the piece covers running end-to-end, property and differential tests before committing, reading REVIEW.md from the main branch, tying review approval to a specific commit, and calling a reliability checklist on trigger conditions.
Why it matters: The Claude Code team moves end-to-end, property and differential tests ahead of the commit and reads review rules from the main branch so pending changes can't edit their own grading criteria.