The Reliable Agent Stack Is Half New, Half 20 Years Old

The reliable agent is not a single invention. It is two halves bolted together, and only one is new — a characterization the reporter infers from Coyle's demonstrated mechanism and Eifrem's vendor-validated three-ontologies pattern, not a direct quote from any single source. Builders are wrapping probabilistic reasoning in a deterministic validation layer, and that layer runs on standards that predate the smartphone: RDFS and OWL for the formal scaffolding, Schema.org, FOAF, and Dublin Core for the shared vocabulary, all already inside the training data of modern LLMs.

At AI Engineer World's Fair, Frank Coyle of UC Berkeley made the case in a 20-minute talk: a Claude agent loop that consults an ontology before answering, using the graph to validate what the model has just produced. His working definition, citing Oxford Semantic Technologies: ontologies describe data structure, naming the classes, properties, and relationships inside a domain of knowledge. The same week, Neo4j CEO Emil Eifrem named three concrete ontologies every enterprise agent stack needs: a business-facing map of organizational concepts, a technical map of all metadata in the data ecosystem, and an execution-trace layer capturing an agent's runtime signals. Neo4j is already shipping ontology-aware agent products, which moves the pattern from academic to vendor-validated.

The portable distinction is the one that matters. Every "reliable agent" pitch that follows can be read as a claim about which seam it has engineered between the probabilistic half and the deterministic half. The seam is the only part of the architecture that is genuinely new. Everything else is plumbing older than most of the people wiring it up.