Онтологии снова в моде: как AI-агенты и Neo4j возрождают семантическую сеть

Engineers building AI agents are massively returning to ontologies in 2026 — structured knowledge graphs from the Semantic Web era — to keep probabilistic language models inside deterministic logical boundaries. The approach that combines neural networks with symbolic rules is called neurosymbolic AI.

Why agents need ontologies

Ontologies give an agent a rigid framework it cannot "drift" away from the facts within: they describe data as a graph — entity types, their properties, and relationships within a specific domain. A large language model excels at probabilistic reasoning, but that same probabilistic nature makes the agent unpredictable; an ontology acts as a logical limiter, keeping the model inside a defined set of rules.

There's no need to build such schemas from scratch — the web has already accumulated mature standards.

• "Neurosymbolic AI" is a pairing of neural networks with symbolic, rule-based systems
• Ready-made web ontologies: Schema.org, FOAF, Dublin Core, RDFS, and OWL (Web Ontology Language)
• Emil Eifrem (CEO of Neo4j) identifies three types of ontologies for agents: business, technical, and execution traces
• OpenLink Software is building an "engineering stack for agents" on Semantic Web technologies
• The topic came up at the AI Engineer World's Fair 2026; a talk by Frank Coyle of UC Berkeley ran 20 minutes

Neo4j's three layers of knowledge

Emil Eifrem, CEO of the graph database Neo4j, breaks ontologies for agents down into three levels. The first is a business ontology, understandable to humans. The second is a technical ontology — that is, corporate metadata. The third is execution traces, the signals an agent leaves behind while it works.

Neo4j embeds ontologies directly into its agentic products, while its competitor OpenLink Software, led by Kingsley Idehen, is assembling a full-fledged agent engineering stack on Semantic Web technologies.

Why the Semantic Web failed before

The first attempts to build a semantic web in the 1990s–2000s failed mainly because ontologies were too expensive to maintain by hand. Standards like RDFS and OWL had existed for decades, but people had to manually describe and update knowledge schemas — and they couldn't keep up.

AI developer Prasenjit Sarkar proposes a solution that was unavailable twenty years ago: let the agent itself maintain the ontologies, updating them right in the course of its work. What once killed the Semantic Web may now become its salvation.

"The beauty of an ontology is that it defines the entity types and relationships through which language gains computable context," —

Kingsley Idehen, founder of OpenLink Software.

What this means

Ontologies are not a step back into the past but a missing layer of control for agents: the more companies entrust language models with real-world actions, the stronger the demand for deterministic boundaries that keep a probabilistic model from making costly mistakes.

Frequently Asked Questions

What is an ontology, in simple terms?

An ontology is a formal knowledge schema in the form of a graph: it lists the types of objects in a domain, their properties, and the relationships between them. For an AI agent, it's a map of permissible facts and relationships that it does not step outside of.

What is neurosymbolic AI?

Neurosymbolic AI is the combination of neural networks with symbolic, rule-based systems. The neural network handles probabilistic language processing, while the symbolic ontology handles deterministic logical constraints.

Why didn't the Semantic Web take off in the 2000s?

According to the discussion participants, the main reason was the cost of manually maintaining ontologies: standards like RDFS and OWL existed, but people had to update the knowledge schemas themselves. The idea in 2026 is to entrust that maintenance to the agents themselves.