AgentLogs

AgentLogs is a dataset of activity related to the GitHub agents functionality: repository metadata, agent tasks, sessions, session logs (messages, tool calls, usage details, etc.), and user records.

This dataset is described in:

Jonan Richards, Kosei Horikawa, Youmei Fan, Yutaro Kashiwa, and Mairieli Wessel (2026), AgentLogs: A Dataset for Opening the Black Box of GitHub's Cloud Agent. arXiv: 2608.29204 (preprint).

	# Records	Size	Table	Content
Repositories − 1.98% with agent tasks	1,812,362 - 35,810	395.6 MB	repositories	Public GitHub repositories with over 10 stars (metadata including name, license, language, stars, forks, timestamps, labels, topics).
Agent tasks − 99.90% found − 97.00% with sessions	307,416 - 307,108 - 298,188	77.1 MB	agent_tasks	Agent assignment on a repository (metadata including name, request, state, creator, timestamps, branch/PR identifiers).
Agent sessions	549,239	225.1 MB	agent_sessions	Agent runs within a task (metadata including model, prompt, outcome, usage, and branch/PR identifier for that session).
Log entries − >99.99% parsed	64,255,174 - 64,254,936	56.0 GB	agent_session_logs	Session log events (including messages, usage details, tool calls for file edits, git, and GitHub issues, PRs, comments, CI).
Users − 99.96% found	33,573 - 33,561	39.7 MB	users	Users related to the agent tasks and sessions (only GitHub id and username).
Total	66,957,764	56.7 GB

See the schema reference for field-level documentation of each table.

Entity relationship diagram for the AgentLogs dataset. Non-foreign key fields are omitted.

Example analysis code and a small sample: risenlab/agentlogs.

Installation

pip install risenlab-agentlogs huggingface_hub datasets

Load

Snapshot (recommended)

Download parquet once, then query with DuckDB, Polars, or the local streaming notebooks. Prefer this for agent_session_logs.

from pathlib import Path
from huggingface_hub import snapshot_download
from agentlogs.schema import assert_dataset_version

dataset_path = Path("data") / "dataset"
snapshot_download(
    repo_id="risenlab/agentlogs",
    repo_type="dataset",
    revision="v0.2",
    local_dir=dataset_path,
)
assert_dataset_version(dataset_path)

load_dataset

Convenient for the smaller tables.

from datasets import load_dataset
from agentlogs.schema import assert_version

assert_version("v0.2")

repositories = load_dataset("risenlab/agentlogs", "repositories", split="train", revision="v0.2")
tasks = load_dataset("risenlab/agentlogs", "agent_tasks", split="train", revision="v0.2")
sessions = load_dataset("risenlab/agentlogs", "agent_sessions", split="train", revision="v0.2")
users = load_dataset("risenlab/agentlogs", "users", split="train", revision="v0.2")

agent_session_logs is large. Prefer a snapshot. Do not stream that config with load_dataset (full rows, including payloads, are downloaded to your machine).

Citation

If you use AgentLogs in an academic publication, please cite this preprint:

@misc{richards2026AgentLogsDatasetOpening,
  title = {{AgentLogs: A Dataset for Opening the Black Box of GitHub's Cloud Agent}},
  author = {Richards, Jonan and Horikawa, Kosei and Fan, Youmei and Kashiwa, Yutaro and Wessel, Mairieli},
  year = 2026,
  month = aug,
  eprint = {2608.29204},
  primaryclass = {cs.SE},
  doi = {10.48550/arXiv.2608.29204},
  archiveprefix = {arXiv},
  url = {https://arxiv.org/abs/2608.29204}
}

License

CC BY 4.0. Suggested attribution:

AgentLogs dataset by RISEN Lab, licensed under CC BY 4.0.

Repository sampling uses GitHub Search (GitHub, Zenodo) © SEART Research Group and Contributors, used under the MIT License.

The GitHub Search seed CSV is not distributed with this dataset. The repositories table is built from that CSV.