Summary
In this AI Engineer conference keynote, Neo4j CEO Emil Eifrem presents what he calls the “thin agents on a smarter shared substrate” pattern — an architectural approach designed to solve the data-wiring problem that emerges when enterprises try to scale AI agents beyond a handful of pilot projects. The core argument is that agents built today are too thick: each one independently discovers data sources, hard-codes connections in prompts and code, and leaves no cross-agent learning. When something changes, every agent breaks.
Eifrem’s proposed solution rests on three pillars: a business-facing ontology that expresses enterprise concepts (customers, accounts, transactions) in human-readable terms; a technical ontology cataloguing the actual databases, Snowflake instances, S3 buckets, and schemas beneath those concepts; and a runtime signal layer where agent execution traces are written back into the same graph, enabling accumulating institutional memory. The graph forms a shared substrate that any agent can query at runtime to discover what data exists, where it lives, whether it can be trusted, and whether access is permitted.
Eifrem grounds the framework in real deployments at a Fortune 20 global bank, a major Bay Area tech platform, and a leading fintech company, and references commentary from the Latent Space podcast reinforcing why markdown-file approaches alone fall short. The talk is particularly useful for platform engineers and AI architects evaluating how to build reusable data infrastructure for multi-agent systems inside large organizations.
📺 Source: AI Engineer · Published July 22, 2026
🏷️ Format: Keynote Launch
