AgentTrove is a large-scale open-source dataset of intelligent agent interaction trajectories released by the OpenThoughts-Agent team. This dataset contains 1,696,847 rows of data, sourced from 219 datasets, covering task domains such as code repair, shell scripting, mathematical problem-solving, programming competitions, and general computing use. All trajectories were collected based on the open-source Harbor agent evaluation and data generation framework and published using the Terminus-2 harness format (a ShareGPT-like dialogue layout).

Data fields:

• Messages: A complete agent interaction trajectory, including roles (user/assistant/tool) and dialogue content in a ShareGPT-like structure.
• original_source: Identifier of the original task source (e.g., swesmith, codeforces, nl2bash, etc.)
• original_teacher: Identifier of the teacher model that generated this trajectory.
• reward: Bonus points for a successful or unsuccessful track completion, typically 1.0 (success) or 0.0 (failure).
• task_id: A unique identifier for a task instance; the format varies depending on the source.
• Other metadata fields: Additional information retained from the original dataset.

Citation

@misc{openthoughts-agent,
author = {Team, OpenThoughts-Agent},
month = Dec,
title = {{OpenThoughts-Agent}},
howpublished = {https://www.open-thoughts.ai/blog/agent},
year = {2025}
}