Terminal-World: Large-Scale Environment Synthesis for Terminal Agents

Terminal agents enable large language models to execute tasks directly in command‑line environments, but their progress is limited by a scarcity of high‑quality training data. Existing data‑generation approaches rely on manually defined seed tasks or local GitHub repositories, leading to narrow task distributions, mismatched environments, and teacher trajectories containing excessive unguided exploration and inefficient operations.

To address these issues, the paper introduces Terminal‑World, a fully automated pipeline for synthesizing terminal‑agent data. The pipeline treats Agent Skills as primitive synthesis units, leveraging the implicit “what to do, when to use, how to execute” information in each skill to jointly generate task instructions, executable environments, and teacher trajectories. To further enlarge the synthesis space, Terminal‑World composes individual skills into skill teams and skill graphs, supporting multi‑role and cross‑domain complex tasks.

Using this pipeline, the authors built 5,723 training environments and trained three model families: Terminal‑World‑8B, Terminal‑World‑14B, and Terminal‑World‑32B. Experiments on six benchmarks show consistent superiority over existing terminal‑agent baselines. Notably, Terminal‑World‑32B, trained with only 1.2 % of the data used for Nemotron‑Terminal, achieves a Pass@1 score of 31.5 on Terminal‑Bench 2.0, surpassing Nemotron‑Terminal‑32B, and reaches a Pass@3 of 43.8.