01
The middle plane
In the pattern worth stealing, analytics, automations, and agents never query tables. They read and act on objects — a Plant with a live efficiency score, an Order with a delivery to schedule, a Customer with an action called notify. Data sources feed the objects from below; models and rules give them behavior; the systems where consequences land sit on the far side.
The plane in the middle is the interface in the strict sense: it is the contract between what the organization knows and what the organization can do. Everything above it inherits its vocabulary, and everything below it disappears behind that vocabulary.
02
Schema is governance
Deciding what counts as an object, which properties it carries, which relations are real, and who may invoke which action — these look like data-modeling choices and are actually governance choices. When access control and audit ride on the object layer, permissions become semantics: the question "what can this agent see and touch?" is answered by the ontology, not by a scattered pile of grants.
This is why the layer is political inside any organization that builds it. The schema encodes whose view of the business wins. An ontology is an org chart for meaning.
03
Why agents force the issue
Language models made unstructured interfaces cheap: point one at raw tables and it will produce plausible answers. But plausible readings of raw data are precisely the unadjudicated-belief failure this notebook keeps circling — and an agent acting from them acts without a consequence boundary.
The ontology is that boundary made of meaning. An agent that can only act through typed objects with named actions is an agent whose behavior is inspectable, permissionable, and auditable by construction. The alternative is an agent whose capabilities are whatever the prompt happened to allow that day.
04
The small version
The pattern scales down further than its enterprise packaging suggests. My own systems keep a typed object layer between ingestion and action: entities carry provenance, confidence, and domain boundaries — the veracity gradient lives on these objects — and no automation acts on anything that lacks a name, an owner, and a permitted set of actions.
You do not need a platform to have an ontology. You need the discipline. The open question I am still testing is whether ontology-first survives contact with fast-moving research code, or whether the schema ossifies faster than the work changes shape.