Neo4j CEO Emil Eifrem argued at GraphSummit that manually wiring data connections for AI agents will fail as enterprises scale from a handful to hundreds of agents. He proposed thin agents operating on an ontology-based semantic layer with three pillars: a business-facing ontology, a technical ontology of all data assets, and execution traces enabling agents to learn from past runs. This approach, already in production at a Fortune 20 bank, a major Bay Area tech platform, and a leading fintech company, replaces repeated manual data discovery with a shared, queryable graph. Eifrem described a typical Fortune 500 company where a team building an account-opening agent wires in customer identity data, then a second team building a fraud detection agent starts from scratch discovering the same sources. When a data source migrates or schema changes, every agent must be rewired, and no execution learning passes between agents. Eifrem stated this pattern has unfolded over “the last call it six to nine months” as large enterprises moved from experimenting with AI agents to deploying them in production. He delivered the prescription: “We need thin agents on a smarter shared substrate.” The argument arrives as OpenAI urges enterprises to shift from tracking token costs to measuring work completed, while Thinking Machines Lab released the open-weight Inkling model, arguing organizations should customize AI on proprietary data.