Study: AI agents increased code volume by 30%, but not software releases — Ars Technica
A Harvard University study found that introducing AI agents for programming increases the volume of code created, but is not accompanied by a statistically significant increase in software releases. As Ars Technica reports, gains at the code-writing stage are offset by longer and more complex human review of the results.
More code and merge requests
Researchers Fiona Chen and James Stratton analyzed aggregated data from the Jellyfish platform: 300 million work events, including commits and code merge requests, as well as data from task management systems. The sample covered more than 700,000 employees at over 700 software development companies from 2021 through March 2026.
After AI agents were introduced, the total number of lines of code at companies increased by an average of 30%, the number of commits by 20%, and the number of code merge requests by 23%. At the same time, the rates of closing Issues and Epics in systems such as Jira, which track the completion of tasks and major product features, did not change in a statistically significant way. The authors also found no changes in the size or complexity of such tasks after AI tools were introduced.
More current news is available on the UA.News Telegram channel Telegram.
Code review became the bottleneck
The average time from submitting a merge request to its integration into the codebase increased by 49% after AI agents were introduced. The share of requests in which reviewers required changes nearly doubled, while the number of comments per request rose by 35%.
In response, the share of employees involved in code review increased by 14%. At the same time, the researchers were unable to link AI use to significant changes in overall employment at these companies after comparing Jellyfish and LinkedIn data.
As of March 2026, 80% of companies in the sample used some AI tools for code review. However, AI agents left 23.3% of all comments during reviews and were responsible for 10.8% of all merge requests, so most of this work remained with humans. According to the authors, 95% of companies in the sample had already introduced AI agents for programming.
