The bottom line: Established web ontologies such as Schema.org and OWL serve as “logical guardrails” for LLM-based agents and are already embedded in their training materials.
UC Berkeley Professor Frank Coyle argues that AI agents need formal ontologies alongside language models to function reliably. He coined the term “neuro-symbolic AI” for this – the combination of probabilistic language models and rule-based logic.
In a 20-minute presentation at the AI Engineer World’s Fair 2026, Frank Coyle, computer scientist and lecturer in Generative AI at UC Berkeley, presented the rediscovery of ontologies for modern agentic systems. While language models deliver excellent probabilistic reasoning performance, they lack, according to Coyle, structural guardrails for reliable agent decision-making – precisely the role of ontologies.
In computer science, an ontology is a description of data structures: classes, properties and relations in a knowledge domain. Coyle defined it pragmatically as “data as graphs”. Established web standards such as Schema.org, FOAF (Friend of a Friend), Dublin Core, and technologies such as RDF (Resource Description Framework Schema) and OWL (Web Ontology Language) are already anchored in LLM training materials, allowing developers to leverage these proven structures directly rather than reinventing them. The company Neo4j employs ontologies in its agent products: CEO Emil Eifrem described three layers – a business-oriented ontology for core concepts, a technical one for metadata across all data sources, and execution traces for runtime signals.
Coyle coined the term “neuro-symbolic AI” for this fusion: neural networks coupled with rule-based, symbolic systems and knowledge graphs. This keeps the language model structurally on course. Kingsley Idehen from OpenLink Software, who is developing an “Agent Engineering Stack” with Semantic Web technologies, emphasizes that ontologies combine the typical language competence of LLMs with semantic contexts – thereby giving language computability. However, maintainability and currency of ontologies remain classic challenges that persist in modern agent engineering practice.
Source: www.latent.space · Published 30 July 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.