Terminal agents are moving from "run a few commands on your machine" toward completing professional work independently. The new bottleneck is training data: most synthetic environments produce short, one-shot trajectories — copy a file, install a dependency, run a test — leaving no room for the skills that actually matter: gathering evidence through exploration, revising strategy from feedback, and turning intermediate results into final deliverables. TerminalHorizon, from Alibaba's AMAP team and the University of Science and Technology of China, targets exactly this gap.
The Blind Spot in Synthetic Data: Environments Do the Work in Advance
The paper is blunt about the problem: environment synthesis often completes or simplifies the key work during construction, "spoiling" the exploration, implementation, and verification the agent was supposed to practice. TerminalHorizon's answer has two stages. First, independent execution probes what key work was omitted when the environment was built, and the resulting evidence guides reconstruction. Second, tasks under the same goal are progressively expanded, introducing new requirements that build on prior results — forcing trajectories to grow cross-stage continuity.
The output: 1,535 executable environments covering 25 professional domains and 463 subdomains, from software engineering and scientific computing to agriculture and public services, organized into 27 behavioral patterns across five families (explore, diagnose, build, verify, coordinate). The "long" in long-horizon is literal: mean interaction length is 52.82 turns (P50 of 40), versus 15.60 for the next-highest method, CLI-Gym.
All 12 Benchmarks Improve
The team supervised-fine-tuned Qwen3.5-27B on 2,852 trajectories to produce TerminalHorizon-27B, which improves on all 12 benchmarks: Terminal-Bench 1.0 from 46.3 to 55.0, 2.0 from 41.6 to 55.8, and 2.1 from 46.82 to 60.67. The largest jump comes on AutomationBench, which measures agentic tasks in real production settings: the partial metric climbs from 26.4 to 49.7 (+23.3 points), and the pass rate goes from 5.5 to 17.2 — more than tripling. General reasoning also rises (AIME 2025 from 86.7 to 93.3), though more modestly.
Two additional experiments stand out. On data scaling, Terminal-Bench 2.1 already reaches 57.30 with 1,000 trajectories and 60.67 with 2,852 — headroom remains. And the recipe transfers: fine-tuning two larger bases (35B-A3B and 122B-A10B) on the same data lifts TB 2.1 from 39.70 to 52.81 and 49.06 to 64.04 respectively.
12-Hour Autonomous Research: Learning Not to Give Up
The most striking part is the RSI (recursive self-improvement) case studies: each case grants a 12-hour research budget to test whether a model can keep experimenting beyond an initial solution, improve its methods from feedback, and retain effective improvements in its deliverable. On a multi-view point-cloud fusion task, the base Qwen3.5-27B stops voluntarily after 12.1 minutes with a final F-score of 53.62; TerminalHorizon-27B keeps iterating until it locks its final code at 2.21 hours, reaching an F-score of 88.12, with the hidden reward rising from 21.93 to 48.96. "Not giving up easily" turns out to be a skill that long trajectories teach.
Caveats and What to Watch
The lane is crowded this year: Qwen's Terminal-Universe reverse-rebuilds old trajectories into 37,300 environments, and FACET builds environments before writing tasks — both focused on where environments come from. TerminalHorizon bets on a different axis: the completeness of the work inside each trajectory. All-12 gains and doubling-level improvements on production tasks suggest "long" is a real signal. The cold water: only the SFT recipe is validated, general-reasoning gains are limited, and every number is self-reported with no third-party replication yet.
For a work led by a mapping-business team, the industrial intent is clear: AMAP's services hold deep reserves of real professional long-horizon tasks — a data moat few others have. While the field competes on environment-synthesis tricks, "who holds real professional workflows" may be the dividing line for the next stage of terminal agents. Paper: https://arxiv.org/abs/2610.13166