One developer reviews 79% of AI agent pull requests, study finds
Researchers at Rochester Institute of Technology analyzed 25,264 agentic pull requests across GitHub repositories with at least 100 stars, covering GitHub Copilot, OpenAI Codex, and Claude Code activity from May through July 2025. The finding that stands out: 78.9% of those pull requests were reviewed by a single developer, who read the code, fixed what needed fixing, and merged it.
How the workflow actually plays out
The pattern is familiar to anyone who has used an AI coding agent. Prompt it, wait for the pull request, open it, edit a few lines, ship it. Count the cases where a reviewer leaves the agent’s output untouched and one-person oversight covers close to nine in ten agentic pull requests. Group review stays rare at every project size, though larger teams spread the work around more often.
The researchers also found that small repositories (one to five contributors) averaged 50.2 agentic pull requests over the three-month window, far more than medium or large teams, driven by a handful of outlier projects running agents hard. The median repository across all sizes opened just one or two agentic pull requests in the entire quarter.
Solo and group review merge at nearly the same rate
Single-reviewer pull requests merged at an 81.2% rate within the study window. Multi-reviewer pull requests merged at 80.3%. The outcome is nearly identical. The difference is what goes into each bucket: solo reviewers tend to handle feature work, while multi-person workflows cluster around bug fixes.
Lead researcher Maliha Noushin Raida noted that even small teams running more than 30 agentic pull requests over the period kept the single-reviewer pattern in place. More agent activity did not produce more distributed review. The review desk stayed one person wide regardless of throughput.
⏱️ Review is still the ceiling
The study identified 25 projects generating agentic pull requests at a pace comparable to a Worklytics benchmark for professional developer output. Every other project came in below that threshold. The constraint is not the agent’s ability to produce code. It is the human hours available to read, test, and judge whether to merge it.
The dataset covers acceptance within the study window only. Reverts, follow-up fixes, and long-term code quality sit outside the scope of the research.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
