Ontologies for AI Agents

[ 01 / THE SEMANTIC GRAPH ]

Meaning lives in relationships.

A graph database represents domain entities as nodes and their relationships as typed edges. An order is more useful to an agent when it is connected to the customer who placed it, the payment that settled it, and the refund that changed its balance.

Each edge carries operational meaning. Following this path establishes ownership, identifies the relevant payment, and exposes the completed refund. Properties on the nodes and edges carry currency, timestamps, source systems, and confidence. Constraints define a valid transition: the sum of refunds for a payment cannot exceed the amount settled.

Validation follows the graph. Starting from the requester, the system traces the relationships required for the decision and returns the connected evidence as one piece of context.

[ 02 / THE DECISION ]

Trace the evidence.

A customer asks, “Refund my last order.” The agent identifies order 1042 and proposes a $75 refund. Before execution, the semantic service traverses the decision subgraph.

The agent receives a decision and the path that produced it. The response explains that the payment has already been refunded in full, cites the relevant records, and prevents a second transaction.

The same pattern supports other domains: an account owns a subscription that grants an entitlement; a shipment contains a package assigned to a route; an article cites a source that supports a claim. Typed relationships establish a consistent interpretation of the connected facts.

[ 03 / A SECOND DOMAIN ]

Dependencies reveal consequences.

An engineering agent is asked to deploy a change to the billing service. The change looks contained in the repository, but the graph connects it to a database migration and the checkout system that depends on the service.

The traversal changes the decision. The agent holds the deployment and returns the missing verification, the migration it applies to, and the dependent service at risk. The same graph can answer a later question—what should be retested after the database changes—without rebuilding the relationships from a document search.

[ 04 / THE AGENT LOOP ]

Supporting the agent with semantics.

The semantic layer sits alongside the agent loop. It assembles the relevant subgraph before a decision and evaluates the proposed state change before a tool executes it.

The graph connects identifiers and events from operational systems while those systems remain authoritative for transactions. A semantic service resolves entities, traverses their relationships, applies constraints, and returns both the result and its provenance. The agent uses that response to proceed, ask a question, or request approval.

[ 05 / THE FIRST IMPLEMENTATION ]

Start with one decision.

Choose a consequential action and model the relationships needed to decide whether it should proceed.

1. Define the decision.What must the system establish before the action is valid?
2. Map the subgraph.Which entities and typed relationships supply that evidence?
3. Add state and provenance.Where did each fact come from, when was it observed, and is it current?
4. Expose the check.Use one interface that returns a decision, the evidence path, and the allowed next step.

The refund graph is small enough to implement and test against real cases. Each additional decision expands the ontology according to demonstrated need.