Not more text. The relationship.
Ontology is the knowledge graph inside AI Suite. A workspace gets a tenant. You register object types, link types, and actions. Agents read the graph and propose changes. A queued action waits for a person.
Ontology is currently in beta. Pricing will be announced later.
A typed graph, scoped to a workspace.
A tenant is the ontology for one workspace. Every object, link, and type belongs to that tenant. The schema has to exist before an agent can propose an object or an edge of that type.
Tenant
One ontology per workspace. API keys and signed-in members only reach tenants in a workspace they belong to.
Registry
Three kinds of type: object, link, and action. Each registration is versioned. The slug is the name agents use.
Objects and links
An object is a typed node with properties. A link is a directional edge: from one object, through a link type, to another.
A policy on every action
auto and deferred run now. queue returns pending and waits for approval. The audit log is append-only.
Types first. Then the graph.
Object types declare properties. Link types declare which object types they may connect. Action types declare a target and a human-in-the-loop policy: auto, queue, or deferred.
Object type
enterprise
v1 · properties
A slug, a description, and the properties the schema requires. Objects of this type cannot be proposed until the type is registered.
Link type
active_deal
contact → enterprise
From types and to types. The edge is directional. The same pair keeps one link of each type.
Action
queue
HITL policy
auto and deferred apply now. queue returns pending. A person approves or rejects. The log does not update or delete.
Typed nodes
Properties are checked against the type schema. Status is active, draft, or archived. An external id makes a sync idempotent.
Typed edges
First-class and directional. The same pair of objects can hold one edge per link type. An object cannot link to itself.
Governed writes
Execute a registered action. Queue policy stores a pending action with the proposal and its evidence. Approve runs the effect. Reject records a reason.
What an agent may say
A read returns a card: grade, flag, confidence, freshness. The grades are underwrite, cite-with-flag, and do-not-cite. Stale objects come back flagged.
Read, traverse, then propose.
Workspace agents and managed agents share the same tools. They list types, query and search objects, fetch one object with its neighbors, and walk a typed path. Propose tools return an approval card. The write applies when a person approves.
Workspace agentsWorkspace UI
Overview, object types, link types, actions, graph, approvals, and the API. Overview shows objects, links, pending approvals, and freshness.
REST
/api/ontology/v1 — tenants, types, objects, links, actions, traverse, graph, subgraph, stats, activity, and export.
MCP
The same operations behind an API key: search, link, execute action, approvals, stats, and the subgraph the agent walked.
Scopes
ontology:read, ontology:write, and ontology:admin. Read sits inside write. Write sits inside admin. Approving a pending action requires admin.
The Impact Suite is built on Ontology.
SundayPyjamas Impact Foundation runs sundaypyjamas.org as one tenant. That public board is the proof, not the definition of the product. The write-up is the case study.
Give the agent a graph it can walk.
Register the types. Let the agent read and propose. Keep the queued actions for a person. Ontology is currently in beta. Pricing will be announced later.
Or email contact@sundaypyjamas.com - a founder replies.