A study conducted by researchers from Harvard University, Fiona Chen and James Stratton, has found that while AI coding agents generate more lines of code, they do not significantly increase overall software output for firms. The study analyzed data from over 700,000 employees at more than 700 software development firms, using analytics from Jellyfish, which tracks coding activities. The researchers noted that human code review processes create a bottleneck that absorbs any efficiency gains from AI-generated code.
The study reports a 30 percent increase in total lines of code generated and a 20 percent rise in total commits after the introduction of AI coding agents. However, the resolution rates for software issues tracked by tools like Jira did not show statistically significant changes. The average review process time increased by 49 percent after AI agents were implemented, with the share of pull requests requiring changes nearly doubling and comments per pull request increasing by 35 percent.
Despite a 14 percent increase in the share of workers performing code reviews, the researchers could not attribute significant employment changes to the introduction of AI. Although 80 percent of firms used some form of AI code review by March 2026, human reviewers were still responsible for the majority of review comments and pull requests. The researchers suggest that as firms gain experience with AI coding agents, the trade-offs between coding speed and review time may improve.
The findings indicate that while AI coding agents can enhance coding speed, they also increase the time and effort required for human review, raising questions about the overall value of implementing such tools in software development.