Agentic AI Needs Ontologies for Guardrails, Says UC Berkeley Expert

In the rapidly evolving world of artificial intelligence, the ability of AI agents to operate reliably and coherently is paramount. Frank P. Coyle, PhD, an educator at UC Berkeley with decades of experience in computer science, recently delved into the critical role of ontologies in achieving this goal during a presentation at the AI Engineer World's Fair.

The Convergence of Agents and Ontologies

Coyle highlighted the current AI landscape, noting that while the computer science degree was once a guaranteed path to employment, the rise of AI has shifted the focus. He emphasized that AI and agents are now the key areas to explore, posing the question of how to effectively navigate this new technological universe.

His core argument centered on the synergy between AI agents, which are probabilistic in nature, and ontologies, which provide formal structure and logic. Coyle explained that agents, conceived as entities that perceive, decide, and act, are the driving force behind many current AI applications. However, their probabilistic nature, while enabling creativity and pattern recognition, also leads to potential issues like hallucinations.

Ontologies, on the other hand, offer a formal specification of a shared conceptualization. Coyle traced their origins back to Aristotle's categories of being and their more recent formalization by philosophers like Quine and Gruber. He described ontologies as providing the necessary "guardrails" for AI agents, ensuring their outputs are grounded in logical constraints and world knowledge.

Building and Utilizing Ontologies

Coyle outlined two primary methods for building ontologies: top-down and bottom-up. The top-down approach involves experts defining entities, properties, and relationships, reminiscent of early expert systems. The bottom-up approach extracts information from data sources, such as documents and customer reviews, using techniques like named entity recognition and LLM mining.

He also stressed the importance of reusing existing ontologies, such as Schema.org, FOAF, Dublin Core, SKOS, and DBPedia (which powers Wikipedia), to avoid reinventing the wheel and to benefit from established knowledge structures.

The presentation also touched upon augmenting ontologies with technologies like RDFS and OWL (Web Ontology Language). Coyle explained how these can enable inference and enforce constraints, such as ensuring that if 'A' is an ancestor of 'B', and 'B' is an ancestor of 'C', then 'A' is an ancestor of 'C'. Functional properties, like 'has father', can also be constrained to ensure logical consistency.

The Role of Loops and Guardrails

Coyle then shifted to the concept of agents and loops, noting that while loops are fundamental to computation, they also present dangers like infinite loops, goal drift, and token cost blow-ups. He argued that ontologies, through RDFS and OWL, can act as the necessary guardrails to keep these loops in check and ensure agents operate within defined boundaries.

He provided a code example illustrating an agent loop, emphasizing how ontologies can be used to validate both the input to a tool call and the output received. This validation process, he explained, helps to ensure the agent's actions are reasonable and aligned with the desired outcomes.

A Philosophy of Making

Coyle concluded with a powerful educational philosophy, quoting Sister Corita Kent: "Nothing is a mistake. There is no win and no fail. There's only make." He encouraged a hands-on, iterative approach to learning and building, emphasizing the importance of engaging one's senses and actively creating rather than passively consuming information.

Ultimately, Coyle's message underscored the critical need for a structured, logical framework like ontologies to complement the probabilistic power of LLMs, leading to the development of more reliable, understandable, and safe AI agents.