"Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments" by Jie Wu , Zhenru Zhang , Beichen Zhang , Xuwu Wang , Yuhui Su , Mouxiang Chen , Peng Wang , Zhihai Wang , Que Shen , Hao Zhou , An Yang , Fei Huang , Yujiu Yang , Dayiheng Liu
TLDR:
Terminal-Universe is a framework developed to enhance training tasks and post-training performance in terminal-based code agents. It reconstructs reusable environments from agent trajectories, allowing for the synthesis of new tasks and interactions. By leveraging the tool-execution history found in trajectories, Terminal-Universe recreates environments where the agents ran, extracting specific tasks and providing execution feedback. The framework reconstructs the original intent task and generates new tasks by replaying file operations and restoring modified files in a partial workspace. Additionally, it scales tasks by mining dependency relations between environments for breadth and extending single-turn queries to multi-round sessions for depth, capturing iterative user feedback. Tested on public terminal agent trajectories, Terminal-Universe significantly improves performance metrics on various benchmarks after supervised fine-tuning.
Read Paper / Blog
1September 5, 2026 54
