Early Adoption of Agentic Coding Tools by GitHub Projects

Asteria Okay, so there's a paper out of Rochester that analyzed twenty-five thousand agentic pull requests across twenty-three hundred GitHub repos, and the headline is kind of wild — everyone's using agents, but almost nobody is actually using agents.

Draco Right. The median repo generates one to two agentic PRs in three months. That's basically nothing.

Asteria It's the product adoption story we keep seeing, right? Widespread shallow adoption masquerading as mainstream use. The tools are there, projects have them, but they're not actually integrated into the workflow yet.

Draco The study is from May, June, July twenty twenty-five, so it's a year-old snapshot. But the mechanism they're pointing at is solid. They looked at three agents — Copilot, Codex, and Claude Code — and tracked how projects actually manage the PRs once they're submitted. And the answer is: one person does it.

Asteria One person reviews and modifies the agent contributions. That's the mode.

Draco Mm-hm.

Asteria Which means the constraint isn't the agent's ability to generate code anymore. It's literally how many PRs one developer can read and decide on. And if that developer is also the one who has to fix the agent's mistakes… you've got a ceiling pretty fast.

Draco That's the thing that jumped out. Most papers focus on PR-level outcomes — does the PR get accepted, what's the quality, does it merge. But this one is asking: at the project level, how do you actually organize oversight? And the answer is… poorly. Or at least, not yet. They're not doing distributed review, they're not building governance structures. It's one person per agent, and that person is a bottleneck.

Asteria Small projects, one to five contributors, show way higher participation and more agentic PR activity per capita than medium or large teams. Why do you think that is?

Draco Probably selection bias. Small projects that adopted agents might be the ones where one person is already wearing all the hats. Larger teams might have more skepticism about agents, or they might be more cautious about letting the agent submit PRs without more ceremony.

Asteria Or they have more institutional inertia. Like, you can't just let an agent submit PRs to a project with thirty people without some kind of review policy. Small team? One person decides, done.

Draco Yeah. And the productivity numbers bear that out. Most projects are way below the industry estimate of thirty-six PRs per participant in three months.

Asteria So the productivity gains are real but they're not evenly distributed. Which suggests it's not just about agent capability — it's about the team that knows how to use it.

Draco Exactly. The authors actually say this directly: the effectiveness depends not only on agent capability but on human oversight and organizational processes.

Asteria Right. So here's my read — this is the infrastructure-readiness story again. Teams that succeed with agents are the ones that have already solved the hard part: they know what code they want, they have good tests, they have clear standards. But most teams aren't there yet. They're still figuring out what the agent should even try to do.

Draco And the single-human oversight model is a symptom of that. If you had robust specs and good tests, you could let the agent run and the tests would tell you whether it worked. You wouldn't need one person to read every PR and second-guess the agent's judgment. But most projects don't have that. So one person becomes the decider and the gatekeeper and the fixer.

Asteria That person burns out real fast.

Draco Or they just stop accepting agent PRs.

Asteria Which we might be seeing already. I mean, the paper is from a year ago. The adoption patterns have probably shifted since then.

Draco Right, and they acknowledge that in the paper — they call it an early snapshot. The landscape in twenty twenty-six is different. But the core finding is probably still solid: adoption is driven by governance and review capacity, not by agent output.

Asteria So the product win isn't in the agent getting smarter. It's in the infrastructure that makes oversight scale. Better specs, better tests, better visibility into what the agent did.

Draco Or tools that help distribute the review burden. Let multiple people sign off without duplicating work.

Asteria Which is… infrastructure work. Boring infrastructure work. The kind that doesn't ship as a feature, it ships as a team practice that someone has to build and maintain.

Draco Yeah. And that's why I think the paper is actually more useful than a lot of the capability studies. It's telling you where the real friction is. Not in the agent's code quality. In the human decision-making bottleneck.

Asteria The study is solid too. Twenty-five thousand PRs across twenty-three hundred repos is a real dataset, not a benchmark or a toy example.

Draco The fact that they went project-level instead of just PR-level is the thing. You can drown in PR-level metrics and miss that most projects are barely using agents at all.

Asteria Okay, I think that's the read. Agentic tools are spreading wide but shallow. The constraint is human oversight and team governance. Small teams are using them more because they have less governance overhead. And most projects are nowhere near the productivity ceiling. Which means there's still a ton of room to grow, but the growth depends on solving the organizational problem, not the capability problem.

Draco Ask me in six months whether adoption has actually deepened or if we're still in the wide-shallow phase.

Asteria Fair bet. This is the kind of study that's going to matter more in hindsight than it does right now.