The Pulse
AI Agents Raise Code Output 30%, but Software Output Gains Are Statistically Unclear
AI coding agents increased lines of code per worker by 30% at adopting firms, while estimated gains in completed software output were small and statistically unclear. The study links the gap to a growing code-review bottleneck, with review
AI.info Team · Written with AI ·
AI coding agents increased lines of code per worker by 30% at adopting firms, while researchers estimated small positive but statistically insignificant changes in completed software output. The findings come from a study by Harvard researchers Fiona Chen and James Stratton, whose current paper version is dated August 4, 2026. The gap points to a constraint beyond code generation: teams still need to review, revise and approve what agents produce.
More commits, small and statistically unclear output estimates
Chen and Stratton analyzed about 300 million work events from 718 firms between January 2021 and March 2026, including GitHub activity, Jira issues and Google Calendar events. Their staggered adoption analysis compared firms’ activity before and after they introduced AI coding assistants or agents. The agent results were larger than those for assistants, which did not significantly change lines of code or pull requests.
After adopting agents, firms saw an estimated 1,494 additional lines of code per worker each month, a 30% increase over the study’s baseline. Commits rose 20%, and pull requests—the proposed changes submitted for team review—increased 23%. Estimates for Jira issues resolved and larger Jira epics completed were small and positive, but not statistically significant.
The researchers define a Jira issue as completed when the associated code reaches production. Their estimates rule out an output increase above 12% of the baseline mean for agents, though they do not prove that the effect is exactly zero. They also found no evidence that teams shifted toward larger or more complex issues after adoption, which could otherwise have made a flat issue count misleading.
Reviews take longer as agent-written code arrives
The study found the sharpest change between submitting a pull request and merging it. Average review time rose 49%, from a baseline of 7.03 days; the share of pull requests receiving a request for changes nearly doubled, and comments per request climbed 35%. The paper found no significant change in pull-request size, suggesting the heavier review was not simply a response to larger submissions.
“For AI agents, this incomplete pass-through reflects a bottleneck from code review: review times increase, a larger share of code updates require revisions, and reviews involve more comments.”
Fiona Chen and James Stratton, the study’s authors, write
Firms responded by drawing more employees into review: the share of workers participating in code reviews rose 14% relative to its baseline. The share of workers writing code did not significantly change. In other words, the study describes added review work alongside higher coding activity, not a simple replacement of human contributors.
AI review tools have not removed the human queue
By March 2026, nearly 80% of firms in the sample had used AI tools for code review. Yet AI generated only 23.3% of review comments, and just 10.8% of pull requests received at least one AI-generated comment. Human reviewers therefore remained involved in most of the review activity measured by the study.
The researchers also found little evidence that AI adoption changed software output or employment. The results apply to the firms and period studied, and the authors’ method estimates changes associated with adoption rather than establishing that every company will see the same effects. For engineering managers, the paper’s concrete warning is that more code and more pull requests can arrive without a statistically clear increase in Jira work reaching production—and with reviews taking longer.