One of the most watched presentations at the 2026 AI Engineer World’s Fair was a 20-minute talk by Frank Coyle, a computer science professor at UC Berkeley who teaches generative AI and LLMs. Drawing on decades of experience, Coyle reintroduced ontologies to today’s AI engineers and argued that while large language models are strong at probabilistic reasoning, agentic systems need “logical guardrails” — by which he meant ontologies — to be truly effective.
What is an ontology?
In computer science an ontology describes the classes, properties and relationships in a domain of knowledge, a definition used by Oxford Semantic Technologies. Coyle summarized it more simply as “data as graphs.” He noted that ontological thinking stretches back to Aristotle and has recurred throughout AI history.
Industry use cases and three ontology types
Neo4j, known for graph databases, is applying ontologies in its agent products. In a keynote at AIEWF, Neo4j CEO Emil Eifrem described three kinds of ontologies to support a “smarter shared substrate” for running agents at scale: a business-facing ontology that describes an organization’s key concepts; a technical ontology that captures “all the metadata of all the data sources and data assets in your enterprise ecosystem”; and execution traces that represent runtime signals produced by agents.
Eifrem contended these layers enable a shift from “thick agents with manually wired data sources” to “thin agents” that run on a shared ontology-based semantic layer.
The Semantic Web and standards
Coyle highlighted web ontologies and standards as especially useful for building agentic systems, mentioning Schema.org, FOAF, Dublin Core and augmenting technologies like RDFS and OWL. One practical advantage is that established web ontologies are already present in LLM training corpora, so developers can prompt for them instead of inventing ontologies from scratch.
As an illustration, Coyle showed a Claude agent loop that used an ontology to validate the LLM’s reasoning after a tool executed.
Neurosymbolic AI: tying neural and symbolic approaches
Coyle labeled the convergence of probabilistic agents and ontologies “neurosymbolic AI”: neural networks linked with symbolic AI, including rule-based systems and knowledge graphs. He argued this approach provides a way to keep LLMs within guardrails and reduce unwanted behaviors.
Practical experience and limitations
Kingsley Idehen of OpenLink Software, who has long worked with ontologies and is building an agent engineering stack that includes an “agent with RDF memory,” emphasized complementary strengths: “The beauty of LLMs is that they are powerful processors of language. The beauty of an ontology is that it defines the types of entities and relationships through which language acquires computable context.” Together, he said, they bring the expressive power of language into computing’s UI/UX layer.
A recurring challenge is maintenance: keeping ontologies updated and relevant is difficult, which partly explains why the 1990s–2000s vision of the Semantic Web did not fully materialize. Developer Prasenjit Sarkar suggested on X that when an agent maintains the ontology as part of its operation — updating definitions when encountering edge cases — the nature of the maintenance problem changes, though it remains hard.
Despite these obstacles, ontologies’ structured nature pairs well with probabilistic LLMs: the models provide flexible language reasoning, while ontologies act as checks and constraints on outputs.
Loops, guardrails and OWL checks
Coyle also discussed loop engineering, noting that loops are longstanding in computer science but can break or “go off the rails.” Here an ontology can serve as a guardrail: one slide described ontologies as “a bounded set of rules around an unbounded loop.”
Near the end of his talk he demonstrated using OWL to check agents: while natural language can be slippery, “an OWL axiom is a rule a machine enforces.” He showed how a reasoner built on an ontology can help keep an LLM on track.
Why this matters now
At AIEWF many speakers emphasized quality control for loop engineering and the need for humans in the loop, so the renewed interest in ontologies aligns with broader concerns about safety and reliability in agentic systems. Traditional web technologies that underpinned the earlier Semantic Web are finding new relevance: structured vocabularies, OWL/RDFS standards and knowledge graphs can help make LLM-driven agents more interoperable and verifiable.
In short, 2026 is seeing a return of software engineering discipline — and with it a revival of web ontologies — as a pragmatic way to add logical structure and guardrails to probabilistic language models.
