Frank Coyle

Frank Coyle is a researcher and educator affiliated with UC Berkeley who presented the talk "Why Agentic Systems Need Ontologies," arguing that Ontology|ontologies and Neurosymbolic AI provide the architectural foundation necessary to make LLM-based agentic AI systems reliable and safe.

Core Thesis: Ontologies as Agent Guardrails

Coyle's central argument is that natural-language prompt engineering alone is insufficient to constrain LLM agent behavior, and that formal ontological structures must be incorporated into agent architectures. He frames this as giving agents "our conceptualization of the universe, our universe, our domains" rather than relying on imprecise English instructions. (59:40)

He advocates for Neurosymbolic AI as the synthesis of these approaches: "neuro-symbolic AI sort of represents a way to keep the LLM on its guardrails." (1:07:29) His position is that the probabilistic, generative nature of LLMs must be complemented by the formal, deterministic constraints that symbolic AI and ontologies provide.

Hallucination as Feature, Not Bug

Coyle takes a notable and somewhat contrarian stance on LLM hallucination, claiming it is "a feature, not a bug, because probabilistic generation is fundamental to how large language models work." (1:24:10) This reframing motivates his architectural prescriptions: rather than trying to eliminate probabilistic behavior from LLMs, the solution is to place formal validators outside the model to catch domain errors before they have effects.

Historical Context: Expert Systems Revisited

Coyle situates current agentic AI within the longer history of AI, arguing that "agentic AI loops are revisiting and extending the world of expert systems and symbolic AI from the 1980s." (3:55:30) Importantly, he rehabilitates the symbolic AI tradition, contending that "expert systems failed not because the symbolic AI approach was wrong but because they could not scale." (1:59:50) LLMs, in his view, solve the scaling problem, while ontologies restore the formal reasoning that expert systems provided.

The Ontology Validator Architecture

Coyle advocates for a specific agent architecture centered on an Ontology Validator operating within the tool-use loop:

• No side effects before validation: "Agents should be designed to have no side effects until their proposed actions are validated against the ontology." (5:05:20)
• Layered validation: Pydantic handles input parameter type validation, while an ontology reasoner validates outputs — "Pydantic should be used to validate parameter types at input and an ontology reasoner should validate results at output in an agent tool-use loop." (5:03:50)
• OWL constraints for domain logic: OWL functional and disjoint properties can enforce constraints — such as an order being refunded only once or a payment going only to a buyer — "that are difficult to express reliably in natural language prompts." (5:18:30)
• Thinking about the validator: "think about the validator as operating with these ontologies about our domain." (4:44:30)

He demonstrated this architecture using a Claude agent as a concrete example: "I want to show you a little example using Claude agent." (3:57:10)

Practical Recommendations

Beyond the architectural philosophy, Coyle offers practical guidance for ontology construction:

• Reuse existing ontologies: He recommends leveraging established taxonomies such as schema.org, FOAF, and Dublin Core rather than building from scratch, citing reuse as a significant advantage. (2:16:00)
• Use Pydantic: "Something that you should be at least taking a look at if you're doing some of this coding is something called Pydantic." (5:02:20)