HEXIS translates skill instructions and tool interfaces into extended finite state machines. The machine keeps progress and intermediate data explicit, while transition conditions decide which operation comes next. Trace-driven refinements are accepted only after static checks and replaying all previously accepted traces. Across four benchmarks and four executors, the paper reports a 16.1-point average success gain over Skill + ReAct; Qwen3.8-27B used 38.4% to 88.9% fewer execution tokens.
HEXIS compiles agent skills into checked state machines
2026-09-25Frontier watchWATCH
HEXIS translates skill instructions and tool interfaces into extended finite state machines. The machine keeps progress and intermediate data explicit, while transition conditions decide which operation comes next. Trace-driven refinements are accepted only after static checks and replaying all previously accepted traces. Across four benchmarks and four executors, the paper reports a 16.1-point average success gain over Skill + ReAct; Qwen3.8-27B used 38.4% to 88.9% fewer execution tokens.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams with procedural agent skills, this offers a middle ground between free-form prompting and hand-written workflow code. Trial it on compliance-heavy or failure-prone skills, but measure authoring overhead and behavior outside the compiled paths before replacing simpler instructions.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | compile skill clauses and tools into extended finite state machines |
|---|---|
| Validation | static checks plus replay of all accepted traces |
| Evaluation | four benchmarks and four executors |
| Success Gain | +16.1 percentage points over Skill + ReAct on average |
| Token Reduction | 38.4-88.9% for Qwen3.8-27B |
| Artifact | the paper names an anonymous ICLR 2027 artifact, but it was not publicly readable during this run |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 2
Verification cost 3
Risk 2