Harvard researchers find AI coding agents shifted the bottleneck to code review
Scored daily by a customisable AI persona to surface the most relevant engineering leadership news.
Harvard research on AI coding agents shifting bottleneck to code review is directly relevant to AI/ML in SDLC, highly actionable.
A Harvard working paper analyzing 718 firms found that AI coding agents increased code output by 30% but did not measurably raise finished work, while code review time ballooned by 49%. The study, based on Jellyfish analytics data covering 68.3 million Jira issues and 200.9 million commits, uses a staggered difference-in-differences design and attributes adoption timing to procurement delays rather than firm growth. Legal AI tools like Harvey's motion-to-dismiss workflow face a similar human-checkpoint bottleneck, yet neither Harvey nor Cooley's redaction agent (developed with Google Cloud) reports review-stage metrics.