AI coding agents now produce code far faster than humans can review it, shifting the core engineering bottleneck from writing to verification. Data from multiple sources (Faros AI, CodeRabbit, GitClear, GitHub) shows AI adoption drives ~4x raw output but only ~12% real productivity gain, while code churn rises 861%, defect rates jump from 9% to 54%, and zero-review merges increase 31%. The author argues review is now the most leveraged skill in software. Key recommendations include tiering review effort by blast radius (not author), requiring PRs to arrive with stated intent and test evidence, running multiple heterogeneous AI reviewers rather than one, treating CI gates as non-negotiable, scrutinizing test changes carefully, and keeping a human accountable for every merge. The appropriate level of human review scales with blast radius: solo developers with no users can lean heavily on AI review, while teams maintaining large systems for many users must keep humans on load-bearing paths. The author uses Claude Code and Codex to triage PR queues, allocating human attention to high-risk changes rather than reading every diff.
