An Ontology for Large Language Model Based Agent Systems
Abstract
As large language model (LLM) based agents pro-liferate, frameworks such as LangChain, AutoGen and CrewAI express similar concepts under different names and abstractions, while existing ontological approaches either focus on the work-flow level or do not jointly cover LLM-specific layers such as memory, sessions and tool use. This paper proposes an OWL 2 ontology for LLM-based agent systems comprising 47 classes, 51 object properties and 79 data properties across six layers (agent core, tool use, memory, session/message flow, model con-figuration and multi-agent orchestration), reusing the OASIS 2 foundational ontology. The design is tested in four independent ways: consistency checking and automatic classification with the HermiT reasoner, answering nine competency questions via SPARQL 1.1, design-quality scanning with OOPS! (zero Critical pitfalls), and a five-scenario fault-injection experiment together with a scalability test up to 120K individuals. Framework independence is further evaluated quantitatively by mapping the live class hierarchies of LangChain, CrewAI and AutoGen to the 37 ontology terms through reflection (coverage: 78–92%).