Repositories running each agent
At the end of each month. A repository leaves a line when it removes that agent.
Loading…
Repositories now start coding agents such as Claude Code, Copilot and Codex from GitHub Actions, where they review, triage and change code with the token and secrets the workflow hands them. GHAgentCI holds the complete histories of the workflow files that do this, mined from 196,253 active GitHub repositories, with every agent invocation located at the step that makes it.
Repositories sampled from SEART on 3 October 2026 (non-forks, at least 300 commits, active in the past year) and mined between 3 and 5 October 2026. Counts include named agents with strong evidence only.
At the end of each month. A repository leaves a line when it removes that agent.
Loading…
Share of mined repositories that ever ran an agent from CI; languages with at least 1,000 repositories.
The explorer has much more: how agents are wired in, the events that start them, the safeguards around them, how long projects keep them, and every repository's workflow histories.
All files are on Zenodo under doi:10.5281/zenodo.24088832, released under CC BY 4.0. Author and committer emails are replaced by a salted hash; names are kept as git records them. Every column is documented in DATASET.md.
| File | One row per | Size |
|---|---|---|
repositories.csv.gz | repository of the sampling frame, agentic or not, including those that could not be mined | 4 MB |
revisions.csv.gz | change to a workflow file whose history is agentic at some point, from its creation to its deletion | 17 MB |
steps.csv.gz | agent invocation: agent, evidence, job and step, action and version, triggers, purpose | 10 MB |
ghaw_manifests.csv.gz | revision of a GitHub Agentic Workflows source or compiled file: engine, tools, permitted outputs, network, sandbox | 2 MB |
content.tar.gz | distinct workflow revision, as content/<sha256>.yml | 808 MB |
frame-2026-10-03.csv.gz | repository of the sampling frame, with SEART metadata (language, stars, creation date) | 7 MB |
replication.tar.gz | the tool, run logs, scripts and validation material to rebuild the dataset and the paper's numbers | 123 MB |
Four tables plus the content store. A file's history is identified by (repository, uid): the uid survives renames and content changes.
repository; a history is (repository, uid).(repository, uid, commit_hash).content_sha256; frame joins repositories on name = repository.Each notebook downloads what it needs from Zenodo and runs as is on Google Colab, or locally with uv run --group notebooks jupyter lab.
Load the tables, see how they join, and reproduce the headline numbers.
When repositories adopt agents, which agents, in which languages, and how many stop.
Which events start agents, who may trigger them, their permissions, and how actions are pinned.
GitHub Agentic Workflows manifests, and reading a workflow's YAML from the content store.
The dataset was built with cai-miner, an open-source tool that parses every revision of every workflow file and classifies agent evidence into strong and weak tiers. It works on local clones and needs no API token.
pip install cai-miner
git clone --filter=blob:limit=64k --no-checkout \
https://github.com/OWNER/NAME.git clone
cai-miner clone -o out/
import pandas as pd
base = "https://zenodo.org/records/24088832/files"
steps = pd.read_csv(f"{base}/steps.csv.gz?download=1")
steps.groupby("agent").repository.nunique()
The accompanying paper is under review at the MSR 2027 Data and Tool Showcase Track. Until it is published, please cite the dataset:
@dataset{ogenrwot2026ghagentci,
author = {Ogenrwot, Daniel and Delgado Cervantes, Jorge Gonzalo and
Kim, Woojoong and Businge, John},
title = {{GHAgentCI}: A Dataset of Coding-Agent Invocations in
{CI} Workflow Histories},
year = {2026},
publisher = {Zenodo},
version = {2026-11-09},
doi = {10.5281/zenodo.24088832},
url = {https://doi.org/10.5281/zenodo.24088832}
}