Scheduling-Structural-Logical Representation for Agent Skills
A new arXiv paper proposes an explicit representation for LLM agent skills, moving beyond text-heavy artifacts like SKILL.md documents. The approach draws on Schank and Abelson's classical work on Memory Organization Packets, Script Theory, and Conceptual Dependency to structure skill knowledge into scheduling, structural, and logical components. This aims to make skill collections easier for machines to acquire and leverage by separating invocation interfaces, execution structure, and side effects from natural language.
Key facts
- arXiv:2604.24026v2
- LLM agents rely on reusable skills combining instructions, control flow, constraints, and tool calls
- Current skills are represented by text-heavy artifacts like SKILL.md documents
- Machine-usable evidence remains embedded in natural-language descriptions
- Proposed representation draws on Memory Organization Packets, Script Theory, and Conceptual Dependency
- Schank and Abelson's classical work on linguistic knowledge representation is referenced
- The representation is scheduling-structural-logical
- Aims to help machines acquire and leverage skill knowledge
Entities
Institutions
- arXiv