HEXIS compiles agent skills into extended finite state machines and beats Skill + ReAct by 16.1 points.

arXiv 2609.30123 argues SKILL.md-style skills force the model to re-derive control flow every run, so it skips or misorders steps. HEXIS keeps skill knowledge as local per-state instructions and moves step order into explicit transitions, cutting execution tokens 38 to 89%, and merges new traces only after static checks and a replay of every previously accepted trace. If one of your skills keeps dropping a step, the fix is to pull its ordered procedure into code the agent calls and keep only the judgment calls in prose.