Agent Skills let AI systems represent open-ended task knowledge in natural language, but composing them is tricky. The skills themselves may be flexible, yet putting them together demands a persistent execution order and clear handoffs of outputs between steps. Without that structure, the composition can break down.
A new arXiv paper, MetaSkill, argues that the solution is selective formalization at skill boundaries. Instead of formalizing the entire skill, only the coordination relations between skills are made explicit. This keeps the natural-language flexibility inside each skill while providing the structure needed for reliable composition.
The authors propose that this boundary-level formalization is enough to support complex task orchestration. The approach is still early, but it points toward a middle ground between fully natural-language agents and fully formal planning systems.