Terminal-World: Automated Pipeline for Scaling Terminal-Agent Environments
A team of researchers has unveiled Terminal-World, an automated system designed to utilize agent skills as the primary method for producing high-quality training data for terminal agents. These terminal agents enhance Large Language Models by enabling them to perform tasks in command-line settings. However, the development has been hindered by a lack of quality training data. Current methods rely on limited sources like human-defined seeds or GitHub repositories, resulting in tasks that are restricted to narrow seed distributions, misaligned environments, and inefficient exploration paths. Terminal-World overcomes these challenges by generating task instructions, environments, and teacher trajectories from agent skills, which detail what to achieve, when to act, and how to proceed. Additionally, it combines skills into teams to expand the synthesis range. This research is detailed in a paper on arXiv, identified as 2605.20876.
Key facts
- Terminal-World is a fully automated pipeline for scaling terminal-agent environments.
- It uses agent skills as the central synthesis primitive.
- Existing approaches bootstrap from partial sources like human-defined seeds or GitHub repositories.
- Terminal-World co-derives task instructions, environments, and teacher trajectories.
- Agent skills encode what to accomplish, when to apply, and how to execute.
- The pipeline composes skills into skill teams to broaden the synthesis space.
- The paper is available on arXiv with ID 2605.20876.
Entities
Institutions
- arXiv