Original title: Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
Jie Wu, Zhenru Zhang, Beichen Zhang, Xuwu Wang +10 · September 2026
Training artificial intelligence agents to code inside computer terminals requires realistic, interactive environments where the code can actually run and give feedback. Today, researchers have many frozen records of past agent actions, called trajectories, but very few working environments to test new ideas. To solve this, researchers built Terminal-Universe. It looks at the history of tools used in an old trajectory and plays file changes in reverse to recover original files. Then, a helper agent fills in missing files and dependencies to recreate a usable workspace. Using these reconstructed environments, the framework rebuilds original tasks and invents new ones. It also connects related codebases and uses a simulated user agentAn AI that simulates user interactions to provide feedback in multi-round sessions. to add multi-step conversations. Tested on public agent trajectoriesRecordings of an AI's actions when writing or executing code., the system created 37.3k working environments. Training a model on this new data boosted its coding test scores by 11.9 points on single-turn tasks and 13.8 points on multi-turn tasks.
Imagine you watch a time-lapse of someone rearranging a messy room. You want to recreate the original setup, but the video only shows the final state. In this work, the recordings of code agents are like that time-lapse, and the researchers figure out how to reverse the changes to restore the room. Then they use the restored room for other tasks.
This matters because training better coding AI requires realistic workspaces, which are hard to create. By recycling old agent recordings, they generate thousands of workspaces, helping AI learn more effectively and improving performance on coding tasks.
The abstract doesn't specify limitations, but the work is applied to public terminal agent trajectories, which may not represent all coding scenarios. It also focuses on terminal-based environments, not graphical interfaces.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments — Jie Wu, Zhenru Zhang, Beichen Zhang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.