Long-horizon agents are expected to complete workflows in which later activities depend on earlier outputs and changes to shared state. Evaluating these agents requires tasks that capture such dependencies, but scaling task requirements while maintaining control over complexity is a known challenge.

The arXiv preprint introduces SkillHorizon, a method for constructing long-horizon agent tasks through skill composition. The abstract frames this as a response to the evaluation bottleneck, though it does not disclose implementation details.

Because the abstract is truncated, the full methodology and experimental results are not available from this source alone. The paper appears to offer a way to systematically build tasks with expandable dependencies, but readers should consult the full text for specifics.