Agentic Much? Adoption of Coding Agents on GitHub

The paper sets out to measure how widely coding agents have been adopted in open-source practice. Its starting observation is that, unlike IDE code completion, which leaves no explicit trace, coding agents such as Claude Code, Codex or Cursor leave abundant traces in software repositories: configuration, rule and guidance files such as AGENTS.md and CLAUDE.md, co-authorship trailers and author metadata in commits, conventional branch names, and pull request labels. Drawing on a set of heuristics the authors maintain for 63 coding agents — 79 author-based, 93 file-based, 20 branch-based and 4 label-based — validated by two-author manual annotation of samples of about 400 flagged files, commits and pull requests, it mines 128,018 active, non-fork GitHub projects with at least 5,000 lines of code and 100 commits, over a study period from 1 January 2025 to 21 February 2026.

It defines a coding agent as an LLM running in a loop with tool access that aims to complete a given goal, and distinguishes it from code completion by the scale of the tasks undertaken and the degree of autonomy — up to submitting a complete pull request for a bug report. The study is organized around six research questions covering overall adoption, its relation to project characteristics, its contexts (organizations, GitHub topics, languages), its evolution over time and across tools, the size of agent-assisted contributions, and the types of those contributions. The authors argue throughout that their figures are more likely to undercount than overcount real use.

Key Points

• As of 21 February 2026, 12.08% of the projects showed coding agent traces at the file level (including agent files listed only in .gitignore) and a further 11.51% only at the commit level, giving a conservative adoption estimate of 22.20% and a high estimate of 28.66%; a footnote reports an updated estimate of 31.63% to 38.23% as of the end of May 2026.
• File-level and commit-level detection overlap only partly: of projects with visible agent files, 64.54% were also detected by commit-level heuristics.
• The paper classifies a project’s commit-level adoption by its AI-assisted commit ratio since adoption into Experimental (below 1%), Limited (1–5%), Consistent (5–20%) and Pervasive (above 20%); among file-level adopters with at least 10 post-adoption commits and some AI-assisted commits, 23.97% were Pervasive.
• File-level adoption is higher for younger projects (26.37% in the youngest decile versus 7.96% for projects over a decade old) and, contrary to the expectation that agents suit greenfield work, also higher for larger projects by lines of code, contributors, commits, issues and pull requests.
• Larger and more established adopters show more Experimental and less Pervasive commit-level adoption, so higher file-level adoption in large projects does not translate into a higher commit ratio.
• Across the repositories of 20 top organizations, file-level adoption is 17.21% against 12.08% overall, a relative increase of about 42%, with wide variation between organizations.
• Adoption is broad across GitHub topics and programming languages, with few topics showing very low adoption; commit ratios vary much less across topics than file-level adoption rates do.
• Cumulative adoption shows inflection points at the end of February 2025, around mid-May 2025, a slowdown in autumn 2025, and a renewed sharp rise at the start of 2026.
• A few tools account for most adoption: GitHub Copilot and Claude Code together account for more than half, and the top five tools — assuming most “Generic” AGENTS.md use is Codex — for more than 80%; a third of adopting projects use more than one tool, the most common pair being Claude Code with Copilot.
• AI-assisted commits are larger than human-authored ones — a median of 31 added lines against 11 — and also delete more lines and touch more files, which the authors read as pointing toward increased churn.
• In a sample of 790 commits co-authored by Claude Code, 35.7% were features and 29.9% bug fixes and only 7.1% chores, so feature commits are about twice as common as in a prior study of human-authored commits, a comparison that rests on that other study’s figures.

Notes

The authors discuss reasons their estimates could be too high — projects that tried agents and stopped, and uncertainty about how much of a detected commit the agent actually wrote — and reasons they could be too low: agents that do not sign commits (Codex, for instance), settings that turn attribution off, developers who commit agent work by hand, agent configuration kept in separate dotfiles repositories, and agents for which they have no heuristic. They conclude an undercount is more likely, citing their finding that half of a sample of projects whose commit ratio fell to zero had added or modified an agent guidance or configuration file after the drop. They also caution that open-source use may differ from industry use, which they expect to be higher given the cost of running agents.

The study counts the guidance files that make agents visible but cannot always attribute them: because AGENTS.md is now read by many agents, the authors file its users under a “Generic” category rather than crediting them to Codex, which originally introduced it (see AGENTS.md). They close with a call for practitioners to document rather than hide their use of coding agents, and for agent providers to keep standardizing while defining mechanisms to identify individual agents. The paper was posted to arXiv as arXiv:2601.18341 in the Software Engineering (cs.SE) category, first submitted on 26 January 2026 and revised on 8 April 2026; its arXiv abstract reports the 22.20% to 28.66% adoption estimate and characterizes the figure as very high for a technology only a few months old. Related empirical work in this wiki includes Adoption and Impact of Command-Line AI Coding Agents.