CodAGE — Coding Agent-generated GitHub Events — is a public dataset of GitHub events attributable to AI coding agents. It is used as the source data for AI-to-AI Code Reviews of GitHub Pull Requests, which draws its population of agent-authored and agent-reviewed pull requests from it.
Contents
The dataset is divided into several event-type subsets. The three drawn on by the study above are CodAGE-PRs, which carries pull-request authorship signatures, and CodAGE-Reviews and CodAGE-ReviewComments, which carry reviewer-side activity.
Its scale is visible in the counts that study reports from the snapshot it used: a broad first attribution pass over CodAGE-PRs yielded 4,563,819 candidate pull requests carrying a provisional agent label, while the two review-side streams held 4,141,598 review events and 8,560,919 review comments.
That first pass includes branch-name evidence. The study above then applied its own two-tier signature framework as a stricter second pass over the same pool — requiring a body signature or a vendor-controlled login, and discarding branch-name-only matches — which is what reduced those 4.5 million candidate PRs to 2,830,284 attributed agent-authored ones. The study notes that both passes are its own, so the comparison between them is an internal consistency check rather than external validation.
Provenance
CodAGE is collected from GHArchive, so its coverage is public GitHub events and excludes private repositories, GitHub Enterprise deployments and other version-control platforms. It is published on Hugging Face at https://huggingface.co/datasets/taher-ghaleb/CodAGE under CC BY 4.0.
The snapshot used by the study above covers 1 January 2024 to 15 April 2026. That study notes two timing points about working from it: despite the window covering the relevant period, many autonomous AI coding agents emerged only in early 2025; and GHArchive attributes lag event data by roughly a quarter, which makes the most recent quarter’s counts lower bounds.
Use
AI-to-AI Code Reviews of GitHub Pull Requests used it to build a dataset of Closed-Loop AI Review — pull requests authored by one AI coding agent and reviewed by another — reporting 248,641 agent-authored PRs that received at least one AI-attributed review, and measuring how reviewer output varied with the authoring agent and with same- versus cross-product pairing.