A new analysis of 25,264 AI-generated pull requests across 2,361 popular GitHub repositories finds the median project files just one or two of them in a three-month window [S1]. That number sits awkwardly next to the industry narrative of coding agents reshaping software development. The question is whether the gap between hype and the commit log tells us the tools are still finding their feet, or that something about how teams work is the real bottleneck.

The numbers behind the noise

The preprint, posted to arXiv on 16 July by researchers at Rochester Institute of Technology, is an early snapshot [S1][P2]. It has not been peer-reviewed. But the sample size gives it weight: 25,264 agentic pull requests drawn from repositories popular enough to surface in GitHub's trending and most-forked lists [S1].

An agentic pull request is one generated and submitted by an AI coding agent, a tool that writes code, opens a PR, and waits for a human to review it. Think of it as a digital junior developer who files their work and asks for feedback.

The headline finding is how thin the activity is. The median repository generates one to two agentic PRs during the three-month observation period [S1]. Intensive adoption, where a project sees dozens or hundreds of agent contributions, remains concentrated in a small subset of projects [S1].

Small teams, big appetite

Here is where the data gets interesting. Repositories with one to five contributors tend to see a greater share of agentic PRs relative to their size, and more agent activity on average, than larger projects [S1]. The teams closest to the code, with the fewest people to convince, are the ones leaning in hardest.

A separate preprint by Robbes, Matricon and colleagues, also on arXiv, examines coding-agent adoption on GitHub from a different angle [P4]. That work, too, is in its early stages with zero citations so far, suggesting the research community is only beginning to grapple with what agent adoption actually looks like at scale.

The RIT study also references an industry-reported estimate of 36 pull requests per participant over three months [S1]. Most projects stay well below that line. A small number blow past it [S1]. The distribution is not a bell curve. It is a long tail.

The single-human bottleneck

When the researchers looked at how humans and agents actually work together, one pattern dominated: a single developer reviews and modifies the agent's contributions [S1]. Multi-human collaboration, where a team collectively reviews agent output, remains uncommon [S1].

This matters because it points to the real constraint. The researchers observe that improving the agents themselves is not enough to ensure their contributions get integrated, the surrounding human and organizational workflows are equally important [S1]. The agent can write the code. The question is whether anyone has the time and process to check it.

What it means

The plain-English read is this: AI coding agents are real, but they are not yet routine. For every project where an agent is filing dozens of PRs a quarter, there are dozens more where it files one or two and the team moves on. The technology works well enough to try. It has not yet worked well enough, or been woven deeply enough into team workflows, to become default infrastructure.

The single-human oversight model tells you why. One developer, often the person who set up the agent, reviews its work. That is sustainable for a small team experimenting with a new tool. It does not scale to a 50-person engineering org without process change, review standards, and someone owning the integration.

What it means for business

For a two-person startup or a solo developer, the findings are encouraging. Small teams are the ones getting the most out of coding agents right now [S1]. If you are a founder writing code with an agent and reviewing its PRs yourself, you are in the cohort where adoption is highest.

For a larger engineering team, the picture is more cautious. The data suggests that simply giving every developer access to a coding agent will not produce a flood of useful PRs. The bottleneck is review capacity and workflow design. Someone has to read what the agent writes, decide if it is correct, and merge or reject it. Without a process for that, the agent's output piles up.

A suburban web agency with five developers might find that one or two agentic PRs per quarter matches their experience: they tried it, it worked for a specific task, and they have not yet built it into their daily flow. The next step is not a better agent. It is a better process for using the one they have.

What we don't know yet

The study is a preprint and has not been peer-reviewed [S1]. The authors themselves describe it as an early snapshot and note that adoption patterns may evolve rapidly [S1].

Several things remain unknown. The study does not report merge or acceptance rates for agentic pull requests, so we cannot tell how many of these 25,264 PRs actually made it into the codebase. The specific tools involved, whether GitHub Copilot, Devin or others, are not named in the abstract [S1]. The dataset covers popular public repositories and may not reflect what is happening inside private enterprise codebases where adoption could look very different.

The 36-PRs-per-participant benchmark is described as an industry-reported estimate with unspecified provenance [S1], so treat it as a rough reference point rather than a validated standard.

The next concrete signal to watch is whether the companion paper by Robbes and colleagues gains traction and corroborates these findings [P4], and whether follow-up studies extend the observation window beyond three months to show if adoption is accelerating or plateauing.

If this kind of evidence-based reporting on AI tools is useful to you, subscribe to keep reading.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.