New Terminal-Universe framework generates 37 thousand realistic terminal environments for agent training

Researchers have introduced Terminal-Universe, a framework that reconstructs complete terminal environments from agent trajectories and ships 37.3 thousand "task-sufficient" workspaces built for post-training and evaluation. Instead of relying on isolated code snippets or single commands, the system restores the full context: file structure, system state, execution history and inter-step dependencies. The paper is available as a pre-print on Papers with Code.

The pipeline ingests trajectories collected from agents performing tasks in real codebases and decomposes them into state-action-result sequences. Each sequence becomes an independent terminal environment that can be re-initialized, executed in, and verified for whether the agent reproduces the original success. The result is a corpus of 37,300 such environments, each containing the minimum required to solve a specific task — no extraneous noise, no external dependencies that weren't captured in the source.

When the researchers ran SFT on the new corpus, they reported an 11.9-point jump in single-round performance on Terminal-Bench 2.1 and a 13.8-point gain in multi-round performance on EvoCode-Bench v2. Both benchmarks test a model's ability to write, execute and debug code inside a live terminal, not merely complete static snippets. The figures come directly from the paper and have not been externally validated.

Until now, most code-agent training corpora have drawn from raw GitHub repositories or synthetic question-answer pairs. The problem: an agent trained on disconnected files never encounters the reality of a terminal — runtime errors, missing dependencies, shifting system state. Terminal-Universe aims to close that gap by supplying environments that run end-to-end without external infrastructure, bringing training closer to deployment conditions.

The paper is a pre-print that has not undergone peer review. The authors did not publish comparison metrics against existing corpora of the same size, nor did they disentangle how much of the improvement stems from corpus size versus reconstruction quality. It is also unclear how diverse the environments are across domains — web, systems, data — or whether they inherit bias toward the tasks prevalent in the original trajectories. Code and data had not been released at the time of publication.