A study of more than 700 software firms found that introducing AI coding agents sharply increases the amount of code produced, but not the amount of software shipped. Researchers at Harvard, using data from Jellyfish covering 300 million work events, measured a 30% rise in lines of code, a 20% rise in commits, and a 23% rise in pull requests after agents were adopted.
That extra code does not translate into more completed features. The resolution rate for issues and epics tracked in tools like Jira showed no statistically significant change. The bottleneck, the authors say, is human code review: average review time grew 49%, the share of pull requests requiring changes nearly doubled, and comments per pull request rose 35%. Firms responded by shifting 14% more workers into review roles.
The study also found no meaningful employment changes attributable to AI. While 80% of firms used some form of AI-assisted review by March 2026, humans still wrote the vast majority of review comments and handled most pull requests. The researchers caution that agents are new and firms may still be learning how to deploy them, but for now the productivity gains from generating code appear to be absorbed by the cost of checking it. The source article is the sole basis for this summary; no other sources were provided.