Formal Skill: A Runtime Abstraction for Efficient LLM Agents
A new research paper introduces Formal Skill, a runtime-native abstraction for LLM agents that replaces informal natural-language skill descriptions with executable state machines and hook policies. The system uses JSON metadata, action schemas, Python executors, and skill-local runtime state to enforce workflow discipline and policy compliance. By moving reusable procedures from prompt text into structured code, Formal Skill aims to improve token efficiency and reliability. The paper is available on arXiv under reference 2605.19604.
Key facts
- arXiv paper 2605.19604 introduces Formal Skill
- Formal Skill is a runtime-native abstraction for LLM agents
- It uses JSON metadata and action schemas
- Reliable Python executors are part of the system
- Hook-governed control logic enforces policies
- Skill-local runtime state is included
- Aims to replace informal Markdown skills and instruction packs
- Targets token efficiency and enforceable control
Entities
Institutions
- arXiv