Original title: Environment Evolution for Terminal Agents
Zhiyuan Fan, Tinghao Yu, Yuanjun Cai, Jiang Zhou +8 · September 2026
Computer programs called terminal agents learn by practicing in interactive virtual setups called environments. As these agents get smarter, basic practice setups become too easy, so the models stop learning. Older methods make new setups by watching the agent fail, but this stops working well as models improve. The researchers built a method called environment evolution. Instead of relying on the agent's current mistakes, it steadily ramps up the difficulty of environments on its own. It uses three paths that make tasks harder, handled by a team of cooperating agents. Tests showed this system consistently creates tougher challenges. When tested in long training runs, it raised the scores of two models by 14.4 and 18.0 percentage points on a test called Terminal-Bench 2.1.
Imagine you're learning to play a game where the challenges get harder as you improve. If the game only makes new levels after you finish the current one, you might get stuck waiting. Here, the training challenges evolve in advance, scheduled to increase steadily, so the learner always has something appropriately difficult to work on.
As AI models get stronger, they need harder challenges to keep improving. This method ensures they don't outgrow their training, leading to more capable agents for real-world tasks like automation and software development.
The abstract doesn't say how well this method generalizes to other types of agents or environments. It is validated only on specific models and the Terminal-Bench 2.1 benchmark.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Environment Evolution for Terminal Agents — Zhiyuan Fan, Tinghao Yu, Yuanjun Cai 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.