AI coding agents are remarkably efficient at generating large volumes of functional code, but a new study suggests those productivity gains evaporate in the code review process.
Harvard University researchers Fiona Chen and James Stratton analyzed data from Jellyfish, a tool that measures engineering team output, across 300 million work events at over 700,000 employees at more than 700 software development firms between 2021 and March 2026. They tracked the introduction of AI coding assistants—which help with auto-completion—and AI coding agents, which autonomously write and submit code based on prompts.
The raw numbers are striking. When firms introduced AI coding agents, they saw a 30 percent increase in total lines of code generated, a 20 percent rise in commits, and a 23 percent increase in pull requests on average. However, these metrics did not translate into increased software output.
According to the researchers, the resolution rate for Issues and Epics (software features tracked in tools like Jira) showed no statistically significant improvement after AI tools were introduced. The study found no meaningful shift in the size or complexity of tracked issues either.
Instead, human code review emerged as a significant bottleneck. The researchers found that the code review process grew substantially longer after AI adoption, with pull requests more likely to require revisions and reviewers leaving more comments. This downstream constraint absorbed the efficiency gains achieved during the coding phase itself.
The study concludes there is “little evidence that firms increase software output or reduce employment” by deploying these tools. Rather than accelerating software development, the introduction of AI coding agents appears to simply shift work from coding to review, leaving overall productivity unchanged.
The findings highlight a gap between the capabilities of AI coding tools and their real-world impact on software development teams. While the tools excel at generating code volume, human oversight requirements neutralize their efficiency benefits at the organizational level.
Key facts
- AI coding agents increased code output by 30% and pull requests by 23% at firms that adopted them
- Despite higher code volume, software feature completion rates (Issues and Epics) did not improve statistically
- Code review processes grew significantly longer, with more revisions and comments required
- The study analyzed 300 million work events across 700+ software firms from 2021-2026
- No significant reduction in employment was found at firms using AI coding tools