HEXIS compiles into extended finite state machines, separating skill knowledge from control flow. Instead of letting the model infer every next action from skill text and interaction history, each state carries local instructions and explicit transition conditions determine what runs next.
Across four benchmarks and four executors, HEXIS beat Skill + ReAct in 15 of 16 settings, improving success by 16.2 percentage points on average. With Qwen3.8-27B, execution tokens fell by 38.4-88.9% across benchmarks.
The result matters because poor procedural compliance is a known failure mode for skill-using agents. Moving control decisions into explicit transitions can make agents follow required steps more reliably and with less repeated reasoning.