Thin agents
What it is
"Thin agents" describes an architecture where individual AI agents carry little or no bespoke data-wiring logic of their own. Instead, they run against a shared, ontology-based semantic layer that already encodes what the enterprise's data means and where to find it: a business-facing ontology (human concepts like Customer or Account), a technical ontology (a catalog of the underlying data sources), and execution traces that let agents learn from each other's past runs.
Why it's emerging now
As organizations spin up dozens or hundreds of LLM agents, each one has typically reimplemented its own connections to databases, APIs, and internal systems — "thick agents" that duplicate integration work and can't share what they learn. Neo4j CEO Emil Eifrem named the "thin agents" pattern in a GraphSummit talk in July 2026, arguing that a shared ontological substrate is what lets an organization scale past a handful of one-off agents without integration cost multiplying with every new agent. The framing landed alongside a broader wave of talks — including UC Berkeley's Frank Coyle — reviving formal ontologies (OWL, RDFS, Schema.org) as infrastructure for agentic systems, which is what pushed it from one vendor's pitch to a trend other independent voices picked up.
How it relates to what you know
It sits next to "Ontology-driven guardrails" in your ledger but solves a different problem. Guardrails use an ontology to validate or constrain what an agent is allowed to do; the thin-agents pattern uses the same kind of ontology as a shared substrate for data discovery and reuse across many agents, so integration work isn't rebuilt per agent. It's also a partial answer to the coordination costs described by "agent sprawl" — a shared substrate is one way to keep sprawl from multiplying wiring work.