Harvard researchers find AI coding agents shifted the bottleneck to code reviewImage: complexdiscovery.com
Cooling

Harvard researchers find AI coding agents shifted the bottleneck to code review

Heat 50 · #25 in AI · Cooling3 reports · 3 sources · First reported Jan 7 · Updated Nov 20

The story

AI coding agents have sharply increased how much code firms produce without a matching increase in finished work, according to an unreviewed job-market paper by Harvard PhD candidates Chen and Stratton. Using Jellyfish data covering 725,938 workers at 718 firms from January 2021 to March 2026, the paper found lines of code up 30%, commits up 20% and pull requests up 23%, while resolved Jira issues were statistically flat. The bottleneck moved to review: time from pull-request submission to merge rose 49%, from 7.03 to 10.5 days, change requests roughly doubled from 13% to about 25%, and comments per pull request rose 35%. AI review tools were used by about 80% of firms by March 2026 but wrote only 23.3% of review comments, and employment stayed flat. A November revision found larger output gains — code up 44% — alongside a 56.1% increase in review time.
Latest development
same post, same pts/cmt as round 163

How it unfolded

  1. same post, same pts/cmt as round 163
  2. same post, same pts/cmt as round 162
  3. ComplexDiscovery notes WSJ/Economist coverage links on Chen's website and an Oct 12 eDiscovery Today post relaying an Ars Technica report
  4. Nov 7/8, 2026 revision (covered by beri.net and devopsness.com, Nov 12-15): 655 firms, 720,542 workers, code +44%, commits +36%, PRs +35.7%, resolved issues +13.7% (borderline), review time +56.1%, review backlog wait +59.9%
  5. AI review tools used by ~80% of firms by Mar 2026 but wrote only 23.3% of review comments
  6. comments per PR +35%
  7. PR submission-to-merge time +49% (7.03 -> 10.5 days)
  8. lines of code +30%, commits +20%, PRs +23%, but Jira issues resolved statistically flat (+0.12 on 3.67 baseline, SE 0.17)
  9. Aug 4, 2026 version (the version the HN-linked October article covers): Jellyfish data, 725,938 workers at 718 firms, Jan 2021-Mar 2026
  10. Chen and Stratton (Harvard PhD candidates), 'Artificial Intelligence in the Firm: Bottlenecks in Software Production', an unreviewed job-market paper

The angle

AI coding agents write 30-44% more code, but shipped work barely moves - because code review is the new bottleneckSuggested format: Short video · for Developers, engineering managers, AI audience

Heat

Now 50 · Peak 50 · 64 recorded checks since Nov 20
Full ranking →