AgentTorrent

A dataset of the context and harness configuration of coding agents in engineered software projects on GitHub. It contains configuration artifacts for six tools, their imported files, agentic planning documents, and commits that credit the tools. The tools are Claude Code, GitHub Copilot, OpenAI Codex, Cursor, OpenCode and Gemini, which includes Antigravity, Google's successor to Gemini CLI.

To cite our dataset, please use:

AgentTorrent: A Dataset of Context and Harness Engineering Artifacts on GitHub.
Sebastian Baltes, Seyedmoein Mohsenimofidi, Levi Böhme, Vahram Nersesyan, Muhammad Auwal Abubakar, Jai Lal Lulla, Matthias Galster, and Christoph Treude.
arXiv:2611.15436, 2026.
[arXiv] [PDF] [BibTeX] [Zenodo]

The earlier dataset that AgentTorrent succeeds is described in:

A Dataset of Agentic AI Coding Tool Configurations.
Matthias Galster, Seyedmoein Mohsenimofidi, Levi Böhme, Jai Lal Lulla, Muhammad Auwal Abubakar, Christoph Treude, and Sebastian Baltes.
Proceedings of the 3rd ACM International Conference on AI-Powered Software (AIware '26), Montreal, Canada, 2026.
[DOI] [PDF] [BibTeX] [Zenodo]

Of the 40,901 engineered software projects in the sample, 10,944 contain a configuration artifact. In these projects, 1,281,628 commits credit one of the six tools and 72,527 credit one of 154 other tools. Compared with the earlier dataset, AgentTorrent adds OpenCode as the sixth tool, as well as plugins and LSP servers as mechanisms. It also adds the files that configuration artifacts import, the commits of other tools, the form in which each commit credits a tool, and agentic planning documents, which spec-driven development frameworks and the planning features of the tools write. Every heuristic cites the vendor documentation or the commits that it rests on. Still, a configuration artifact does not show that a tool is used, and a credit in a commit does not show who wrote the change. Until AgentTorrent 1.0.0 was renumbered from 2.0.0, this site offered its files with the dataset_version 2.0.0. The README of the release defines every column, and the changelog lists every change.

Dashboard

Tool Adoption

Mechanisms

Visualizations

Interactive charts of which tools and mechanisms the repositories configure, how the tools co-occur, and when the repositories created their first configuration artifacts.

Configuration Mechanisms per Tool

Repositories with each mechanism, by the tool attributed to each configuration artifact. Files attributed to several tools count under each tool. Shared files such as AGENTS.md count in no tool column. The last column counts repositories with only shared configuration artifacts (No specific tool). Percentages use the repository count per tool. Column totals can exceed the overall repository count because repositories can configure several tools.

Tool Usage by Programming Language

The main programming languages of the repositories that configure each tool. No specific tool denotes the repositories with only shared configuration artifacts, such as AGENTS.md, .agents/skills or a root .mcp.json. Percentages are relative to the repository count per tool. Column totals can exceed the overall repository count because a repository can configure several tools.

Tool Co-Occurrence

Repositories that configure both tools in a pair. Percentages use the repository count of the less common tool.

Reference Network Between Context Files

Context files that only reference another file, by source and target category. Other covers files such as README.md and CONTRIBUTING.md.

Tool Usage Combinations (Exclusive Intersections)

Each bar counts the repositories that configure exactly its combination of tools, so each repository appears in one bar.

Distribution of Configuration Artifacts per Repository

Counts of context files, skills, and subagents per repository. For each mechanism, percentages use repositories with at least one configuration artifact of that mechanism.

Creation Order Sequences

Context file categories in the order that their earliest files were created in each repository. Creation dates follow the original commit, including before a file was renamed into a category. Curly braces group categories with earliest files created on the same day.

Creation of the Earliest Configuration Artifacts

The cumulative number of repositories by the month in which their earliest file of each kind was created, for five kinds of context file, for skills and subagents, and for planning documents. A file is dated by the commit that created it under whatever name it had then. Files created before 2023 are left out, because they were written for another purpose and only later renamed into a context file or named in a configuration, such as a development document that a repository renamed to claude.md and then to AGENTS.md. The Copilot Instructions line counts .github/copilot-instructions.md and .github/instructions/*.instructions.md.

Repositories

Repositories in which at least one configuration artifact was detected.

AI Commits

Commits that credit any tool, detected through authors, trailers, and messages.

Configuration Artifacts

Browse configuration artifacts by mechanism, or use the planning document and import tabs. Planning documents and imports are not configuration artifacts.

Downloads

Download the files of the dataset, which is licensed under CC BY 4.0 and archived on Zenodo.

Usage Example

The pandas example explores which tools engineered software projects configure, whether agents co-author their own instructions, and how often configuration imports reach no file. The pipeline on Zenodo contains further analysis scripts.

.py
import pandas as pd

pd.set_option("display.width", 120)
pd.set_option("display.max_columns", 10)

# Run this script in the directory of the downloaded release, after extracting repos.csv.7z and
# commits.csv.7z with `7z x <archive>`. The script reads only the CSVs. The files themselves are
# in repos_data.7z, which contains the configuration artifacts and the planning documents, and in
# imported_files.7z, which contains the files that the configuration artifacts import.

# repos.csv lists the GitHub repositories in the sampling frame with their metadata and a label of
# whether each hosts an engineered software project ("true", "false" or "unclear"). Read the label
# as text, because pandas turns a column of only "true" and "false" into booleans. Only the
# configured repositories, which contain a configuration artifact, have a scan date and the TRUE
# and FALSE flags of the tools and mechanisms.
repos = pd.read_csv("repos.csv", dtype={"engineered_project": str})
engineered_repos = repos[repos["engineered_project"] == "true"]
configured_repos = repos[repos["scanned_at"].notna()]

# Which tools do the engineered software projects configure, and does the choice depend on the
# main language of a project?
tools = ["claude", "copilot", "cursor", "gemini", "codex", "opencode", "no_specific_tool"]
uses_tool = configured_repos[tools].eq(True)
print(f"{len(configured_repos):,} of {len(engineered_repos):,} engineered projects are configured")
print(uses_tool.mean().sort_values(ascending=False))
languages = configured_repos["mainLanguage"].value_counts().index[:5]
print(uses_tool.groupby(configured_repos["mainLanguage"]).mean().loc[languages].round(2))

# Do agents co-author their own instructions? context_files.csv has one row per context file, such
# as CLAUDE.md or AGENTS.md. first_commit_ai_created says whether the commit that created the file
# credits one of the six tools, and first_commit_ai_tool names the tools that it credits. Several
# tools read AGENTS.md, so which tools do the commits that created an AGENTS.md credit?
context_files = pd.read_csv("context_files.csv")
names = context_files["context_file"].value_counts().index[:4]
top_files = context_files[context_files["context_file"].isin(names)]
print(top_files.groupby("context_file")["first_commit_ai_created"].apply(lambda s: s.eq(True).mean()))
agents_md = context_files[context_files["context_file"] == "AGENTS.md"]
print(agents_md["first_commit_ai_tool"].str.split(",").explode().value_counts())

# How often does an import reach no file, and what do the other imports reach? imports.csv has one
# row per import in a configuration artifact, such as @docs/style.md in a CLAUDE.md. target_status
# is found for an import that reaches a file of the repository that the tool reads, and otherwise
# names the reason, such as missing. For a found file, target_category names its mechanism. The
# depth is 1 for an import in the configuration artifact itself and 2 or more for an import in a
# file that it imports.
imports = pd.read_csv("imports.csv")
print(imports["target_status"].value_counts(normalize=True))
print(imports.loc[imports["target_status"] == "found", "target_category"].value_counts())
print(imports["depth"].value_counts().sort_index())

# Which spec-driven development frameworks do the projects use, and since when?
# planning-documents.csv lists the specs, plans and task lists that these frameworks and the
# planning features of the tools write. Its framework column names the framework, or the tool for a
# planning feature of a tool.
planning = pd.read_csv("planning-documents.csv")
planning["created_at"] = pd.to_datetime(planning["created_at"], utc=True)
print(planning.groupby("framework").agg(repositories=("repo_name", "nunique"),
                                        first_document=("created_at", "min"))
      .sort_values("repositories", ascending=False).head(8))

# How many commits credit each tool per month, and how do they credit it? commits.csv lists the
# commits that credit one of the six tools, and other_commits.csv those that credit another tool.
# attributions gives the form for each tool: account for a commit by the tool's own account,
# reviewer for the co-author trailer of a committed review suggestion, standard for a line that
# the tool writes itself, and convention for any other line that names the tool.
commits = pd.read_csv("commits.csv", usecols=["repo_name", "commit_timestamp", "ai_tool",
                                              "attributions"])
commits["month"] = pd.to_datetime(commits["commit_timestamp"], utc=True).dt.strftime("%Y-%m")
per_tool = commits.assign(tool=commits["ai_tool"].str.split(",")).explode("tool")
print(per_tool.groupby(["month", "tool"]).size().unstack(fill_value=0).tail(6))
print(commits["attributions"].str.split(";").explode().value_counts().head(10))

# How many skills come with scripts? Each of the other mechanisms has a CSV of its own, such as
# skills.csv, hooks.csv and mcp.csv, with the creation date and the number of commits of each file.
# In skills.csv, scripts says whether the folder of a skill has a scripts directory.
skills = pd.read_csv("skills.csv")
print(f"{skills['scripts'].eq(True).mean():.0%} of the skills have a scripts directory")