Agent skills are meant to package reusable knowledge and instructions, but in practice agents still have to figure out each time how to apply them and which operation should come next. That coupling between task reasoning and control decisions makes the execution of prescribed steps less direct than it could be.
The arXiv paper introduces HEXIS, a method that compiles skills into extended finite state machines. By making the control structure explicit, HEXIS removes the need to repeatedly infer the next action from the skill's content, letting the state machine drive the flow instead.
Because the source is a single abstract, the claims are limited to this proposed design and its intended benefit: a cleaner separation between what to do and how to sequence it. The paper does not provide experimental results or comparisons in the available text, so the practical gains remain to be demonstrated.