At 655 firms, AI coding agents added 36% more pull requests and 56% longer reviews
Finished tasks rose 13.7%, a result the Harvard authors call marginally insignificant. Engineering headcount showed no detectable change.
Fiona Chen and James Stratton of Harvard studied 720,542 workers at 655 software firms, using records from Jellyfish, an engineering analytics platform, from January 2021 to August 2026. They compared firms that adopted AI coding tools earlier with firms that adopted later, and measured the change twelve months after adoption.
Coding assistants such as GitHub Copilot and Cursor came first. After a firm adopted them, pull requests (code changes submitted for review) per worker rose 20.9% and commits 16.8%. Adopting coding agents such as Claude Code added a further 35.7% more pull requests and 36.0% more commits. Resolved Jira issues, the firms’ finished tasks, changed by −1.5% after assistants and +13.7% after agents. The authors call the agent result marginally insignificant and rule out rises above 30.7%.
Review slowed after agents arrived. A pull request took 1.99 days longer to reach a merge, on a baseline of 3.56 days. Review comments per pull request rose 38.6%. About 63% of the extra pull requests were merged, against 86% before. Assistants had no significant effect on any review measure.
Engineering headcount showed no significant change. For agents the authors rule out effects outside −15.6% to +10.5%.
What it means
In these firms, AI agents produced code faster than people could review it, and the extra code did not clearly become more finished work. Most firms already used AI review tools (79.1%), yet humans still wrote most reviews and comments. The employment estimates are too wide to say jobs are safe or at risk.
Caveats
- Firms chose when to adopt, so this is not a randomised trial. Larger firms adopted earlier. The authors control for firm size and expect any remaining bias to overstate the effects.
- Jellyfish is a commercial vendor and its clients opted in, so the data is proprietary and cannot be checked independently. The authors say Jellyfish reviewed the manuscript only for leaked proprietary information, not for content.
- This is a job-market paper, a draft the authors have updated since January 2026. Its lines-of-code and finished-task estimates are marginally insignificant, and few firms have more than twelve months of agent data.
- Use of AI through personal licences is not recorded, so the authors treat take-up as a lower bound. The agent effects are on top of assistants.