AI engineers revive ontologies to impose guardrails on agentic systems

A talk at the recent AI Engineer World’s Fair 2026 by Frank Coyle, a UC Berkeley computer science professor, reintroduced ontologies as a means to impose "logical guardrails" on agentic systems driven by large language models. Coyle defined ontologies as "data as graphs" and traced their lineage from Aristotle through decades of AI research.

Coyle argued that while LLMs excel at probabilistic reasoning, agentic workflows require deterministic boundaries to prevent unbounded loops and reasoning drift. He described the convergence of probabilistic agents with ontologies as "neurosymbolic AI," where neural networks interface with symbolic rule systems and knowledge graphs to keep agents within specified constraints.

Neo4j’s CEO Emil Eifrem outlined three ontology types for scaling agentic systems: a business-facing ontology describing core organizational concepts; a technical ontology capturing metadata across enterprise data assets; and execution traces capturing runtime signals from agents. Eifrem characterized this as moving from "thick agents with manually wired data sources" to "thin agents" operating atop a shared semantic layer.

Coyle highlighted established web ontologies such as Schema.org, FOAF, Dublin Core, RDFS, and OWL as already present in LLM training data, enabling developers to prompt for and reuse them rather than rebuild ontologies from scratch. He demonstrated using OWL axioms as enforceable rules to validate agent outputs and keep loops bounded.

Kingsley Idehen of OpenLink Software, a longtime ontology practitioner, described ontologies as providing computable context that gives language-based systems structure. He argued that combining LLMs’ language processing with ontologies yields a more robust agent engineering stack, including systems with RDF-based memory.

Commentary at the event acknowledged the historical maintenance challenges that limited the Semantic Web’s adoption in the 1990s and 2000s. Some practitioners now suggest that agents themselves could help maintain ontologies by updating definitions when encountering edge cases, though the problem remains difficult.