Ten agent platforms will give you ten taxonomies: assistants, tools, treasuries, indexers, copilots. Ask the only question a verifier cares about — who is accountable? — and there are exactly three answers.
Day 5 of 21 · The missing layer
We took W3C PROV-O's trichotomy as law. Every smart agent is a prov:Person, a prov:Organization, or a prov:SoftwareAgent. Nothing else, ever.
A team is an organization (a derived type: Team ⊑ Organization). An indexer is a service. A workspace coordinator is a service. And a treasury is a service — its own smart account, chartered under the person or organization whose money it holds. A person's agent never pays. It authorizes; the treasury transacts, under a mandate the person signed. Roles live in the profile and the on-chain type record. The class never grows.
This is the first place the ontology carries weight in this series, and it will carry weight every week after. It is not documentation. It is a T-box the code binds to by IRI, and a build gate fails when code invents a term the T-box does not declare.
Why the discipline pays for itself:
Accountability has a shape. A person acts for themselves. An organization acts through people who hold custody of it. A service acts only under delegation from a person or organization — it has no standing of its own. Every grant reads against that shape, and a verifier can tell a person's custody from an org's custody from a service's session — including a treasury's.
Names carry the class, not the capability. Our typed suffixes — .me, .org, .team, .svc, .treasury, .registry — name the derived agent type and are checked against the on-chain record, failing closed on mismatch. They never encode capability, provider or version; those change without the agent becoming a different agent. ANS's a2a://name.capability.provider.v1.certified puts all of that in the name. A name should be stable; everything volatile should be a claim.
A person is never also their organization. The shortcut — "the org is just David's account" — collapses two classes onto one address, and every custody, stewardship and treasury gate downstream loses the ability to tell them apart. We enforce the distinction with a build check on our own demo data, because we made the mistake once and everything went green for the wrong reason.
Relationships say what they are NOT. A treasury is charteredUnder the agent that holds it. The property's own definition says: never authority; the child holds its own custody; a resolver may follow it, a verifier may never read it as permission. That sentence lives in the ontology, not in a prompt or an app table — so every gate reads the same one.
The line: three classes is not a limitation of the model. It is the shape accountability actually has — and the ontology is where that shape is written down once.
Honest limit: the trichotomy, typed names and the binding gate are live. The ontology's coverage of derived service roles is still growing as new domains arrive.
Question: Can your agent model let a verifier distinguish "an organization acted" from "a person acted for an organization" — and where is that distinction written down?
#AIAgents #Provenance #PROVO #Ontology #AgentIdentity
Previously: Day 4 — Credentials rotate. Next: Day 6 — Delegation is the artefact (article).