Terminal agents stop training on short commands: ~3K long trajectories lift Terminal-Bench 2.1 from 46.82 to 60.67

Terminal-agent training data no longer centers on short command demonstrations but on long-horizon trajectories averaging 52.8 turns, letting a model explore, diagnose, implement, and verify across one sustained goal. Earlier terminal-task synthesis methods widened task coverage, yet their trajectories stayed relatively short on average and showed behavioral patterns only over a few steps, offering limited experience in sustained problem solving. Under the same teacher protocol and settings, 300 tasks were randomly sampled from each of several open-source terminal-task synthesis methods; these trajectories average 52.8 interaction turns with a 90th percentile of 98 turns, substantially exceeding the compared methods.

Those long-horizon trajectories fine-tune a 27B-parameter model, which then outperforms its base model on all 12 benchmarks spanning coding, real-world production settings, and general reasoning. The Qwen3.5-27B base model scores 46.3, 41.6, and 46.82 on Terminal-Bench 1.0, 2.0, and 2.1. The fine-tuned TerminalHorizon-27B reaches 55.0, 55.8, and 60.67 on those three, gains of 8.7, 14.2, and 13.85 points; on AutomationBench it adds 11.7 points in Pass Rate and 23.3 in Partial score, and VitaBench improves by 11.3 points.

The long-horizon trajectories come from a fully automated synthesis engine that first uses independent execution to find key work completed or simplified during construction, then rebuilds the environment around that evidence. When construction already performs the analysis, implementation, or verification, the agent skips those steps during solving and the intended behavioral patterns never appear. An independent challenger uninvolved in construction attempts each task using only the public instruction; its trajectory diagnoses which work was omitted, and reconstruction revises the initial state, object relationships, implementation requirements, or tool feedback while updating the instruction, reference solution, and verifier.

Under the same goal, tasks are progressively expanded with new requirements that build on prior results, so behavioral patterns continue across stages. Harder tasks or more interaction turns do not necessarily extend those patterns or strengthen their connections. Expansion uses three interaction designs to tie added requirements to evidence, artifacts, and states encountered or produced during the task, so later work depends on earlier progress; an accepted candidate becomes the parent task for the next expansion.

The engine targets long-horizon professional tasks in terminal environments, covering 25 professional domains, with 1,535 task environments and 2,852 trajectories as training data. Evaluation spans 12 benchmarks plus four 12-hour autonomous research cases, each a single research run scored by the final submission on a hidden test set. A reader interested in long-horizon data recipes can follow the task-construction and trajectory-filtering procedures.

The paper reports that scaling from 1000 to roughly 3k training trajectories raises Terminal-Bench 2.1 to 60.67 and states no sign of performance saturation, without giving results at larger scale. Each of the four autonomous research cases is one research run, and its hidden-set score is an optimization score rather than a task success rate. Trajectory filtering and benchmark decontamination are described in appendices whose individual judgments this report did not re-check.