Harvard paper: AI coding agents shifted the bottleneck to review
Original title: Harvard researchers find AI coding agents shifted the bottleneck to code review
Across 718 firms, code added rose about 30 percent while resolved Jira issues did not measurably increase.
Fiona Chen and James Stratton, Harvard Ph.D. candidates, used Jellyfish and Revelio Labs data on 725,938 workers from January 2021 to March 2026. At firms that already had coding assistants, commits rose 20 percent and pull requests 23 percent, while review took 49 percent longer. Finished issues moved by 0.12 per worker-month, indistinguishable from zero. The August 4, 2026 version is a job-market paper with no sign of completed peer review.
Unconfirmed: The paper is unreviewed, and identification rests on the authors’ claim that procurement timing was unrelated to growth.
- Lines added rose by 1,494 per worker-month from a 5,017 baseline.
- Tables list 68.3 million Jira issues and 49.1 million pull requests.
- Only 213 of 718 firms supplied calendar data.
- GPT-4o labeling of a sample found no shift toward larger issues.
- Interviewees put tool approval at about a day or six to 12 months.
- Organizations
- Harvard UniversityJellyfishRevelio Labs
- AI models
- GPT-4o
From the original article
13 min read at complexdiscovery.com
Editor’s Note: Engineers at firms that adopted AI coding agents added about 30 percent more code, yet finished work did not measurably rise and getting code through review took 49 percent longer, according to a working paper by Harvard Ph. D. candidates Fiona Chen and James Stratton covering 718 firms.
By judahmeek · complexdiscovery.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.