Scaling Agentic Data for Long-Horizon Terminal Intelligence

Shun Zou^1,2*
Ziyu Ma^2*
Yi Zou^1,2*
Yong Wang^2†‡
Lin Chen¹
Zehui Chen¹
Guanhua Chen³
Xiangxiang Chu²
Feng Zhao^1‡

¹University of Science and Technology of China
²AMAP, Alibaba Group
³Southern University of Science and Technology

^*Equal contribution. Work done during internship at AMAP, Alibaba Group. ^†Project lead ^‡Corresponding authors

1,535Executable environments

27Behavioral patterns

25Professional domains

12 / 12Benchmarks improved

Motivation

Terminal commands are only part of autonomous work. Our long-horizon trajectories connect exploration, evidence-based judgment, strategy revision, and verification into a complete problem-solving process.

Short, isolated terminal commands contrasted with a long-horizon workflow that traces dependencies, tests hypotheses, revises plans, tunes methods, and cross-validates results.

Method

Two stages turn professional problems into executable tasks that elicit—and entangle—long-horizon behavioral patterns.

Professional domains

25 domains463 subdomains

From software and scientific computing to education, agriculture, and public services.

Illustrated professional-domain overview spanning education, agriculture, healthcare, engineering, scientific computing, finance and infrastructure.

Agentic behavioral patterns

Exploration, diagnosis, construction, verification, and coordination within one sustained effort.

The paper's taxonomy of 27 behavioral patterns, organized into exploration, diagnosis, construction, verification, and coordination.

27 patterns Five families

Environment showcase

Build software, train models, and solve problems beyond code.

Six synthetic task environments. Previews use original requirements, source excerpts, and input data—not agent rollouts or full simulators.

Trajectory statistics

Longer interactions, with sustained chains of actions and terminal feedback.

52.82Mean interaction turns

40P50

98P90

Shared teacher protocol.
300 sampled tasks per compared method; 2,852 TerminalHorizon trajectories.

Experimental results

Training on 2,852 trajectories improves Qwen3.5-27B across all 12 benchmarks.

Scores (%). Base: Qwen3.5-27B. Gains in percentage points. TB 2.0 base score: official model card. AutomationBench reports two metrics.

Performance at 0, 1,000, 2,000, and 2,852 training trajectories: 46.82, 57.30, 58.43, and 60.67 percent. — Training-data scaling
Terminal-Bench 2.1

Grouped bars comparing Base and TerminalHorizon mean pass@1 (%), with a zero baseline: 27B, 46.82 and 60.67; 35B-A3B, 39.70 and 52.81; 122B-A10B, 49.06 and 64.04. — Model-size comparison
Terminal-Bench 2.1

Long-Horizon Task Applications

Four RSI-Exam cases. The research process, the final result, and the difference between models.

12hAutonomous research budget
per model, per case

BibTeX

@misc{zou2026scalingagenticdatalonghorizon,
  title = {Scaling Agentic Data for Long-Horizon Terminal Intelligence},
  author = {Shun Zou and Ziyu Ma and Yi Zou and Yong Wang and Lin Chen and Zehui Chen and Guanhua Chen and Xiangxiang Chu and Feng Zhao},
  year = {2026},
  eprint = {2610.13166},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2610.13166}
}