News · 2026-10-13

TerminalHorizon reports better agent results by rebuilding tasks around the work they require

TerminalHorizon’s researchers released 1,535 task environments and reported improved agent performance after training on work designed to retain its investigative difficulty. Their central terminal benchmark rose from 46.8% to 60.7% completion under the same testing harness. The contribution is a way to construct better practice for long tasks, rather than proof that an agent can improve itself indefinitely.

Key facts

• The release contains 1,535 environments and 2,852 training trajectories.
• The paper belongs to the October 7–9 submission slate reviewed in the October 13 research run.
• A fine-tuned Qwen3.5-27B achieved 60.7% on Terminal-Bench 2.1, compared with 46.8% for the base model.
• The primary source is Scaling Agentic Data for Long-Horizon Terminal Intelligence, with a project page and release collection.

A training task can accidentally erase the skill it is meant to teach. Suppose a developer should inspect logs, trace a dependency, discover the cause of a failure and verify a repair. If a task author preselects the important log lines and explains the cause in the prompt, the learner practices applying a known fix instead. The resulting exercise can look realistic while omitting the expensive reasoning. TerminalHorizon treats that mismatch as a construction problem, rather than simply asking for longer examples.

The TerminalHorizon team calls its approach “challenge-guided construction,” emphasizing how evidence from an attempted solution reshapes the environment. An agent independently tries a candidate task. The construction process examines whether important work has disappeared, then rebuilds the task so the missing analysis or implementation remains necessary. The analogy is a practical examination whose instructor checks that students must actually diagnose the broken engine, rather than finding the diagnosis helpfully attached to the engine cover.

The next step expands an accepted task while preserving its goal. Requirements can broaden responsibility, deepen a mechanism, introduce dependencies, add evidence or demand a better method. Later requirements depend on earlier artifacts or state changes. That dependence is what makes a task meaningfully long: decisions and outputs accumulate consequences. A large number of conversational turns alone is not enough if each turn is just another isolated question with no relationship to what happened before it.

Candidate environments, reference solutions and success checks are updated together. The authors say they filter trajectories for validity, progress grounded in the environment and useful multi-turn training behavior. That is a stronger claim than “we collected long traces,” because long traces can also contain repetition or aimless exploration. The project page and release collection provide the surrounding artifacts for examining how the suite is organized.

The reported training is supervised fine-tuning: a Qwen3.5-27B model learns from the accepted trajectories. It is not an online service continuously learning from every live user task. The authors report gains across 12 benchmarks covering coding, productivity-oriented agent tasks and general reasoning. Their main base-versus-trained comparisons use the same Terminus2 harness, which reduces one common ambiguity: attributing an improvement to model training when the surrounding tools or retry policy actually changed.

The most legible result is the terminal completion rise from 46.8% to 60.7%, roughly fourteen percentage points. It shows an improvement on a named benchmark and testing setup. It does not establish how often the trained model would finish a particular employer’s work, or how much human supervision it would need. The lesson on task-length evaluation explains why tasks, budgets and success definitions must be carried alongside any summary percentage.

The source record does not provide a verified weight-file download size or a stated minimum graphics-memory requirement for the trained checkpoint. Those deployment facts cannot be recovered from a parameter count. The fine-tuning explanation is relevant to the reported training change, while the agent-harness lesson explains the code surrounding the model. Neither a public collection nor a better score automatically makes the release a supported workplace product.

The strongest counterargument is that synthetic environments and automated checks can still embody their authors’ assumptions. Preserving more work is useful only if it is the right work for the intended users. The authors also report twelve-hour case studies of recursive improvement, but those bounded examples cannot establish open-ended autonomous progress. No independent reproduction or substantive paper-specific community discussion appears in the supplied record. A recorded collection upvote count measures attention rather than trustworthiness.

TerminalHorizon matters because it moves the conversation from more training text to better task structure. Useful practice must keep the learner’s uncertainty, dependencies and need for verification intact. For teams building evaluations, the practical question is whether a task still requires discovery after it has been packaged into a convenient environment. The paper offers a concrete construction method and promising author-run results. Its limit is equally concrete: benchmark improvement after supervised training remains evidence about that suite, not a substitute for testing real professional workflows.

Primary source, verified: read the paper → (arXiv 2610.13166)

Ground Truth. (2026, October 13). TerminalHorizon reports better agent results by rebuilding tasks around the work they require. Ground Truth. https://groundtruth.day/news/terminalhorizon-training-preserves-professional-work.html

@misc{groundtruth:terminalhorizon-training-preserves-professional-work,
  title  = {TerminalHorizon reports better agent results by rebuilding tasks around the work they require},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/terminalhorizon-training-preserves-professional-work.html}
}

Topics: agents · coding · training · synthetic-data · evaluation · research