Frank Coyle on Ontology: Logic Outside the Probabilistic Agent (AI column)
Watch the twenty-minute report by Frank Coyle.Why Agentic Systems Need OntologiesAI Engineer World's Fair 2026. I’ve had a complicated relationship with the word “ontology” for a long time: in conversations about architecture, it was almost a stop word. Too often, this terminology has been used by people who are very far from practice: instead of a working contract, it has been an attempt to classify the whole world first.
That seems to be changing now. Issues of taxonomy, ontology, and knowledge work descended to the level of engineering primitives and began to help larger models be more useful. Coyle’s report is about this transition: not about the philosophy of existence, but about the specific place of the formal domain model within the agent cycle.
Frank Coyle is a lecturer in the School of Information at UC Berkeley; among other things, he teaches a course on knowledge representation for intelligent applications. The main thesis of his speech is simple: probabilistic reasoning should be left inside the LLM, and formal logic should be brought out. The model interprets an unstructured query well, gathers context, and suggests the next action. But she does not have to guess every time what the order is, who has the right to receive payment and what conditions are permissible for delivery. This is already knowledge of a specific domain.
In a simplified form, taxonomy answers the question of what types of objects we have. Ontology adds relationships and rules: the customer places the order, the order contains the goods, the return relates to the order, the support employee and the buyer are different roles. On top of this, you can set limits and draw new facts.
In the report, this idea turns into two gateways around calling the tool: Before calling, Pydantic checks the query form: types, mandatory fields and value ranges. The result is checked with the domain model: are roles, connections, states and cardinality permissible? It is only after testing that the change must enter the real world; the questionable result returns to the cycle or is transmitted to the person.
Coyle reveals very down-to-earth mistakes: refunding money on one order, paying a support officer instead of a buyer, and shipping status. probably shipped where the system expects only paid, shipped or refunded. It is possible to describe such invariants in paragraphs in the system prompt, but the prompt remains a request for the model. A formal rule becomes a common machine-testable contract for different agents and tools.
For me, this is where the word "ontology" ceases to be a stop word. It's not about building the ultimate model for the whole company. The local ontology of the operation is sufficient: what entities are involved, what transitions are allowed, and what must be true before an irreversible action. In this sense, it is next to the API schema, the policy engine. (policy engine)A state machine and database constraints, not next to a beautiful architectural picture.
There is, indeed, an important caveat. In the report OWL (Web Ontology Language) Sometimes it looks like a ready-made validator of business rules, but the semantics of OWL works in an open world: the absence of a fact does not mean its falsity, but rather it is a lie. FunctionalProperty Two values can lead to the conclusion that the values represent one object, rather than the usual validation error. Closed checks such as “no more than one return” or “only one of three states” usually require a separate layer: SHACL, code, database constraints, policy engine, or a combination thereof. And, of course, idempotence and transactional guarantees ontology also does not replace.
But this does not change the main thesis of the report. On the contrary, it makes it more practical: it is not a term in itself that is useful, but an explicit executable domain model that can be versioned, tested, assigned its owner and put before the action of an agent.
The bottom line for me is that big models don’t become more useful when we hope they’ve finally learned the rules of business. They become more useful when the system stops forcing them to guess the rules at every turn. The probabilistic model proposes an action, and the deterministic layer decides whether it is permissible to perform it.
#AI #Agents #Architecture #Engineering #KnowledgeGraphs #Data