News

AI Coding Agents Are Already Spreading Across GitHub, Study Finds

• By John K. Waters
• May 27, 2026

A new large-scale study of GitHub projects suggests that AI coding agents have moved quickly from developer experiment to everyday software engineering practice.

The study, titled “Agentic Much? Adoption of Coding Agents on GitHub,” analyzed 128,018 software projects and found an estimated coding-agent adoption rate of 22.20% to 28.66%. The authors described that level as “very high for a technology only a few months old,” and said adoption was increasing.

The paper was written by Romain Robbes, Théo Matricon, Thomas Degueule, Andre Hora, and Stefano Zacchiroli. It was submitted to arXiv in January and revised in April.

Coding agents are different from earlier AI coding assistants, such as code-completion tools. Instead of merely suggesting snippets, systems such as Cursor, Claude Code, and OpenAI Codex can operate with more autonomy, including generating pull requests from a developer’s task description, according to the study.

The researchers said this shift is important because agentic systems leave clearer traces in software repositories than traditional AI coding assistants. Those traces include co-authored commits and pull requests, providing researchers with a way to measure adoption in real-world projects rather than relying solely on surveys or vendor claims.

The findings indicate that coding agents are not limited to hobby projects or early adopters. The authors found adoption across the “entire spectrum of project maturity,” including established organizations, diverse programming languages, and different project topics.

That breadth makes the report notable for development teams, tool vendors, and engineering managers trying to understand how quickly agentic coding is entering mainstream workflows.

The study also found that agent-assisted commits differ from human-only commits. At the commit level, the authors reported that commits assisted by coding agents were larger than those authored solely by human developers and contained a larger proportion of features and bug fixes.

That finding could cut in more than one direction. Larger commits may suggest that agents can help developers complete more substantial tasks. But larger changes can also be harder to review, test, and maintain, especially when teams lack policies for identifying, labeling, or auditing AI-assisted work.

The report does not claim that coding agents improve software quality or developer productivity. Instead, it calls for further investigation into how developers use these systems in practice. The authors said the findings “highlight the need for further investigation into the practical use of coding agents.”

The study arrives as large software companies and AI vendors are trying to turn coding agents into standard developer tools.

Microsoft has embedded agentic capabilities into GitHub Copilot, including features that let developers assign work to an agent that runs in the background and creates a draft pull request. Microsoft is reportedly using AI internally to reduce developer toil across large codebases, while also tracking metrics such as pull requests completed and hours saved.

The broader market is moving in the same direction. Claude Code, OpenAI Codex, Cursor, and open-source agent projects have become part of a fast-moving shift in developer culture, while also raising concerns about cost, reliability, and unintended behavior.

For software teams, the practical question is no longer whether AI-generated code exists in repositories. The new question is how to manage it.

Organizations may need clearer rules for when agents can open pull requests, what level of human review is required, how AI-assisted commits should be documented, and whether security checks should be stricter when an autonomous tool makes changes across multiple files.

The study’s results also point to a likely change in software engineering research. If coding agents leave identifiable traces in repositories, researchers can begin measuring adoption patterns, project outcomes, review behavior, and long-term maintenance costs with more precision than was possible for earlier code-completion systems.

For now, the report’s main contribution is narrower but important: it provides early empirical evidence that coding agents are already present in a significant share of open-source development.

That evidence gives engineering leaders a more grounded starting point for decisions about AI coding policies. The adoption of coding agents is no longer just a vendor roadmap or a developer anecdote. It is beginning to show up in the public record of software development itself.