HEXIS: Compiling Skills into Extended Finite State Machines

Source facts

HEXIS compiles reusable agent skills into extended finite state machines that separate task knowledge from control flow. The compiler maps skill clauses and tool interfaces to states, local instructions, data bindings, and explicit transitions; updates are accepted only after static checks and replay of current and previously accepted traces. Across four benchmarks and four executors, the preprint reports an average success improvement of 16.1 percentage points over Skill + ReAct and lower execution-token use for Qwen3.8-27B.

Analysis

The proposal moves repeated control decisions out of free-form agent reasoning and into inspectable state machines. Replay gates and explicit transitions provide a concrete mitigation path for omitted or misordered prescribed steps.

Uncertainty/Doubt

• This is an arXiv preprint. The benchmark selection, executor coverage, replay safety, and reported token reductions were not independently reproduced here.