Small teams are the heaviest users of AI coding agents – Help Net Security

Small teams are the heaviest users of AI coding agents

The pull request arrives with the tests already run and the description already written, the work of an agent that handled the whole thing on its own. Somebody still has to read it. On GitHub that somebody is usually one developer sitting alone with the diff, and the rest of the project never sees the code

Maliha Noushin Raida and Daqing Hou at Rochester Institute of Technology sorted 25,264 agentic pull requests by who reviewed them and who committed to them. The repositories all carry at least 100 stars, the agents are the ones most developers have already met, GitHub Copilot and OpenAI Codex and Claude Code, and the window runs from May through July 2025

One developer does the reviewing and the fixing

Most agentic pull requests pass through the hands of a single developer, who reads the agent’s code, fixes what needs fixing, and merges it. The setup covers 78.9% of the pull requests in the dataset

It plays out the way plenty of maintainers already work. Prompt the agent, wait for the pull request, open it, edit a few lines, ship it. Count the cases where someone reviews the agent’s work and leaves it as is, and one-person oversight covers close to nine in ten

Group review stays rare at every size of project. Small teams lean hardest on the solo routine, and bigger teams spread the work around more often without displacing it

Most repos run one or two a quarter

The median repository opened one or two agentic pull requests in three months. That is the whole quarter. For most projects on GitHub the agent turns up now and then and goes quiet again, and only a sliver of contributors ever touches its work

The heavy usage belongs to the tiny teams. Repositories with one to five contributors averaged 50.2 agentic pull requests apiece, many times what medium and large teams managed, on the strength of a few outliers running agents hard. Line the small projects up and the one in the middle looks like everybody else

More output, same one reviewer

Raida went back to the small teams that cleared 30 agentic pull requests in the window and checked who was reviewing them

“Even among the most active small repositories (i.e., small teams with more than 30 agentic PRs), the majority of agentic PRs continued to follow a single-reviewer workflow,” Raida told Help Net Security. “This suggests that increased agentic activity did not necessarily lead to more distributed review practices.”

The agent scales up. The review desk stays one person wide

Review is the ceiling

The catch is review. Code review already eats a large slice of the workweek for most developers, and every agent pull request still has to be read, tested, and judged worth merging. That work sits with a human, and there are only so many hours in it

The limit shows up in the numbers. Twenty-five projects in the sample generated agent pull requests at the clip a working developer ships code, going by a Worklytics benchmark of professional output. Everybody else came in under it

Solo and group review end the same way

Pull requests handled by one person get merged at close to the same rate as those handled by several

“The merge rate (i.e., the percentage of PRs merged within the timeframe described in the paper) was very similar between the two collaboration patterns: 81.2% for single-reviewer pull requests and 80.3% for multi-reviewer/committer pull requests,” Raida said

The two groups do different jobs. According to Raida, “the most common type of PR among single-reviewer workflows was feature-related, whereas fix-related PRs were the most common among multi-human workflows.” Solo maintainers point their agents at new functionality. A crowd shows up for the repairs

Those numbers cover acceptance inside the study window and stop there. Reverts, follow-up bug fixes, and how well merged agent code holds up over time sit outside the dataset

The one-person review desk

The pattern is ordinary practice, familiar to anyone who has merged their own Copilot output on a quiet afternoon, and it holds up under load. Small teams running dozens of agent pull requests kept one person in the loop the whole time. Few projects have pushed hard enough against that ceiling to find out where it gives

Must read:

• 20 open-source cybersecurity tools to keep your team ready for anything
• GitHub CISO on security strategy and collaborating with the open-source community

Subscribe to the Help Net Security ad-free monthly newsletter to stay informed on the essential open-

More about

Source: www.helpnetsecurity.com