Scaling Agentic Data for Long-Horizon Terminal Intelligence

What it claims

Environment synthesis for terminal-agent training often does the hard part ahead of time: while building a task environment, the generator completes or simplifies the work the agent was supposed to explore, implement and verify. Trajectories collected in those environments look long but never exercise the behaviours that matter for sustained autonomous work. These are accumulating evidence, forming judgments from it, adjusting strategy on feedback, and turning intermediate results into deliverables. TerminalHorizon is a fully automated engine built to keep that work in the task. Fine-tuning Qwen3.5-27B on 1.5K environments from it produces TerminalHorizon-27B, which the authors report gains on all 12 benchmarks they test.

How it works

The abstract gives these steps:

1. Seed from real work. Tasks start from real-world professional work, not synthetic prompts.
2. Audit by independent execution. A separate run executes the task to detect key work that environment construction left out or pre-completed.
3. Rebuild from evidence. That execution evidence guides reconstruction of the environment, so the exploration, implementation and verification steps fall back to the agent.
4. Expand progressively. The task is extended under the same goal with new requirements that build on prior results. Long horizons come from chained, dependent stages, not from padding a single task.
5. Train. Supervised fine-tuning only, on Qwen3.5-27B. No RL stage is mentioned.

The useful idea, if you build your own agent data pipeline, is step 2. It checks whether a synthesized environment still requires the work, rather than checking only whether the task is solvable.

How to use it

The project site is https://zs1314.github.io/TerminalHorizon. To pull the paper text into an agent, the page gives:

hf papers read 2610.13166

If your CLI does not have that command, install the latest version with:

curl -LsSf https://hf.co/cli/install.sh | bash

The page does not say whether the weights, the 1.5K-environment dataset or the engine code are released, or under what licence. Check the project site before planning to adopt any of them.

Limits

• Abstract only. The fetched page contains just the abstract, with no figures or tables behind "substantial performance gains".
• No benchmark detail. The 12 benchmarks are not named, and no scores or baselines are given. Even the comparison against base Qwen3.5-27B is unstated.
• No cost figures. Neither synthesis cost nor training cost is reported.
• Anecdotal RSI evidence. The RSI results are described as case studies, each with a 12-hour budget. They are not a benchmark.