A new arXiv paper introduces SkillJuror, a framework for evaluating how the organization of agent skills changes runtime behavior.

The authors separate what a skill says from how it is structured. They compare progressive disclosure, where a concise root file points to supporting resources, against a normalized flat baseline.

The result is relevant for agent builders because procedural knowledge is not just content. File structure, disclosure, and retrieval paths can change how agents act during tasks.