Independent Research study
Harvard study of 718 firms: AI agents raise code volume but not shipped work, review becomes the bottleneck
Staggered difference-in-differences across ~300M work events from Jellyfish (Jan 2021–Mar 2026). AI coding agents increased lines, commits and PRs, but resolved Jira issues and epics did not change significantly. Review time rose, more PRs needed changes, and more workers did reviews. No significant employment effect was found.
The numbers (as reported)
| Firms | 718 |
|---|---|
| Work events | ~300 million |
| Lines of code after agent adoption | +30% |
| Commits | +20% |
| Pull requests | +23% |
| Review time (PR submitted → merged) | +49% |
| Comments per PR | +35% |
| Share of workers doing code review | +14% |
| Resolved Jira issues/epics | No statistically significant change |
Facts
- Company / subject
- 718 firms (Jellyfish data)
- Tools
- AI coding agents (various), AI coding assistants (various)
- Codebase
- —
- Work type
- Research study
- Stack
- —
- Published
- 2026-08-04
- Last checked
- 2026-10-17
Caveats
Current version dated 4 Aug 2026 (first version 7 Jan 2026). Data from Jellyfish, an engineering analytics vendor; authors state conclusions are their own. Percentages as reported by Ars Technica from the paper.
Source
Chen & Stratton, Harvard (job market paper); coverage in Ars Technica
Also: https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/