A working paper by Harvard researchers Fiona Chen and James Stratton, covered by Ars Technica in October 2026, finds that AI coding agents increase the code engineers write by about 30% but do not increase the software firms release. The study draws on Jellyfish engineering data from more than 700 firms, and the time from opening a pull request to merging it rose 49%.
What the study found
According to Ars Technica's report, the authors analysed about 300 million work events from more than 700,000 employees at software firms between 2021 and March 2026. After firms adopted AI coding agents, total lines of code rose 30%, commits rose 20% and pull requests rose 23%.
The resolution rate for Jira issues and epics, which the authors use as a proxy for finished features, did not change in a statistically significant way. The researchers also found no shift in the size or complexity of those issues, and they could not attribute significant employment changes to AI.
Where the extra code went
The gains moved into review. The average time between a pull request being submitted and merged rose 49%. The share of pull requests with changes requested nearly doubled, comments per pull request rose 35%, and the share of workers doing code review rose 14%.
AI review tools were common by March 2026, with 80% of the firms using one. Even so, AI agents wrote only 23.3% of review comments, so people still did most of the reviewing. The authors conclude that realising the benefits of AI adoption requires investment in downstream review capacity, not simply more code output.
What it means for engineering teams
The study is observational and covers firms that use Jellyfish, so it describes a pattern and does not prove a cause. Its practical point still holds for any team that gives engineers coding agents: the limit moves from writing code to checking it. A team that doubles generation without changing how it reviews will see longer queues, not faster releases.
That makes review capacity something to plan for. Teams can keep pull requests small, write requirements and tests before the code, and decide who has the authority to approve agent-written changes.
What it means when you hire engineers
If review is the constraint, the engineers who matter most are the ones who can read a large change quickly and tell whether it is correct. That is a senior skill, and a CV that lists AI tools does not show it. Our how we vet process tests it directly, in a real codebase with the candidate's own agents switched on.
To see how we place engineers who work this way, read about our AI-native engineers.
Sources
