TerminalHorizon: 1,500 synthesized environments lift a 27B terminal agent on all 12 benchmarks

Short agent trajectories show commands and tool calls. They do not show the thing that makes a terminal agent useful on a real job: gathering evidence by exploring, changing strategy when feedback says so, and turning intermediate results into a deliverable. Shun Zou, Ziyu Ma and colleagues argue the usual environment synthesis pipelines actively destroy that signal, because the synthesizer completes or simplifies the key work in advance and the agent never has to explore, implement or verify anything. TerminalHorizon is their fix. Starting from real professional work, an independent execution pass identifies which key work the constructed environment omitted, and that evidence drives environment reconstruction. Then tasks are progressively expanded under the same goal with new requirements that build on prior results, so one environment produces a long-horizon chain instead of a one-shot.

With the engine they built 1,500 task environments, fine-tuned Qwen3.5-27B on the resulting trajectories and got TerminalHorizon-27B. The claim is gains on all 12 benchmarks spanning coding agent tasks, agentic tasks in production settings, and general reasoning. The part that will get quoted is the RSI case studies: given a 12 hour research budget, the model keeps experimenting past its first working solution and improves its own results instead of stopping at the first green test.

This is the third environment paper in two days after SoftwareGym and ProgramGym, and they agree on the thesis that environments, not model weights, are the scarce asset for agent training, while disagreeing on how to get them. TerminalHorizon's distinctive move is using an independent run to audit the synthesizer for shortcuts, which is a verifier applied to the data pipeline rather than to the agent. It sits at 132 upvotes on the Hugging Face board.

Links: https://arxiv.org/abs/2610.13166 , https://huggingface.co/papers/2610.13166