skip to content
updated view markdown ->

Generated from agentregistrar/spec at v1.2.0 (tag v1.2.0). Do not edit this page; edit the spec and re-run pnpm sync:spec.

Identity vs Vendor Scope Boundaries

This document is the rulebook the identity protocol must satisfy. It states what the identity protocol owns, what the vendor owns, and the checklist any future change must pass before a new protocol field is added.


1. The Boundary in One Sentence

The protocol proves who signed up, on whose behalf, under which ToS, when, and with what cryptographic proof. Everything about what the vendor sells and how to use its API belongs to the vendor, documented in a post-signup context guide — never in discovery or the signed receipt.


2. In-Scope / Out-of-Scope

2.1 In scope (broker / protocol)

Area Examples
Agent sub-identity registration and human binding identity_register, agent ↔ verified human owner
Master Agent Delegation Agreement and delegated scopes delegation.scopes, tos:accept
Agent Identity JWT assertions and JWKS verification Contract A, issuer-hosted JWKS
MCP OAuth 2.1, MCP tools, agent reachability <agent>@<domain> inbox, inbox tools
Relying Party discovery and brokered signup agent-signup.json, POST endpoints.agent_signup
Delegated RP ToS acceptance terms, tos_acceptances, hash verification
Identity-focused signup ledger and traceability who/whom/which ToS/when/proof
Abuse reporting and universal revocation Contract D.2, kill switch
Post-signup context pointer (identity-adjacent) opaque context.url + context.sha256
Relying Party identity provider account, origin/audience verification, RP signing keys, broker-issued attestation, RP-signed receipts, authenticated abuse identity
RP trust policy trusted RP-broker list, per-RP allow/warn/require/block, assurance level display

2.2 Out of scope (vendor responsibility)

Area Examples
Product catalogs, SKUs, pricing tiers, free-tier quotas instance_catalog, product specs
Resource provisioning APIs and lifecycle /api/provision, instances, endpoints
Resource credentials vendor API keys for resources
Vendor API/SDK usage and operational runbooks usage docs, error catalogs
Legacy email/OTP signup flows implemented by the vendor /api/signup/email, /api/signup/verify-otp
Vendor admin dashboards and instance "nuke" mechanics admin UI, instance actions

2.3 The one-way door

  • Vendor capabilities MUST NOT be represented as protocol surface in discovery, in the signed receipt (Contract E), in broker storage, in the broker UI, in MCP instructions, or in the implementation registry.
  • Vendor capabilities remain available to the agent through the vendor's post-signup context guide.
  • A document does not become conformant by hidden vendor data under extensions.

3. Naming Collision: "Provision"

"Provision" appears in two unrelated meanings. Conflating them is the root cause of much of the scope creep.

Phrase Meaning Verdict
"Auto-provision an agent sub-identity" Broker-side identity_register / MCP context creating the agent's identity and binding it to a verified human. In scope. Not scope creep.
"Provision a compute instance" A cloud provider's /api/provision creating a resource (toaster, hamster wheel, bucket). Vendor scope. Must never appear in protocol surface.

Rule: if "provision" refers to an agent identity, it is identity protocol. If it refers to a resource, it is vendor scope.


4. History

The boundary rules in this document were applied during the identity-only v1.0.0 remediation of the protocol. Internal records of that remediation — the audit findings, the finding-to-change map, and the private repository paths involved — live in the private repository and are not part of this public edition.


5. Do Not Over-Correct

Removing vendor capabilities from the protocol MUST NOT remove these in-scope features:

Concept Verdict
identity_register / "auto-provision" of an agent sub-identity In scope. Different meaning of "provision".
Broker inbox, inbox_wait_for_email, OTP reception, human email forwarding In scope (reachability).
Vendor OTP endpoints (/api/signup/email, /api/signup/verify-otp) Vendor responsibility; may stay as a vendor feature, but leaves protocol discovery.
POST /api/broker/report-abuse, kill switch, revocation In scope.
Agent/human identity, assertion/JWKS, MCP OAuth, master ToS + RP ToS acceptance In scope.
"Signed-Up Services & Proof Ledger" as an identity ledger (which agent signed up with which vendor, which ToS, when, human email shared) In scope. Only the vendor resource columns are not.
Vendor ToS text mentioning appliances/SKUs Harmless; it is the vendor's own terms.
Vendor admin dashboards, nuke actions, storefronts Vendor-owned; not protocol.
provider_account_id in the receipt In scope as an opaque reference; not the start of a resource inventory.

6. Relying Party Identity Is Identity

v1.2.0 adds a second identity to the protocol: the relying party's own broker-issued identity. It belongs here for the same reason the agent's identity does — it proves who the counterparty is, not what they sell.

In scope as protocol (RP identity):

  • the provider account that owns an origin: registration, origin/audience proof, and RP signing keys with proof of possession;
  • the broker-issued RP attestation and its publication in the optional rp_attestation field;
  • the provider API that issues and renews attestations (Contract H) and the broker metadata that advertises JWKS/status/abuse endpoints (Contract I);
  • RP-signed receipts and authenticated receipt ingestion;
  • the RP abuse-report signature and the human-visible trust/distrust policy.

Out of scope as vendor responsibility (even when the RP is verified):

  • product catalogs, SKUs, pricing, free-tier quotas, and provisioning APIs;
  • resource inventory, credential custody, instance lifecycle, and support tooling;
  • a provider account's commercial terms, billing checkout, or pricing model;
  • any per-provider capability list in discovery or the registry.

A provider account is not an agent-facing discovery field. It is the RP's private relationship with its broker. Agents see the public attestation, the assurance level, and the trust decision; they never need an account with the RP's broker, and assertion verification never requires one.

Verification is identity, not endorsement: a verified RP is a registered, origin-controlled entity, not a safe, solvent, or well-behaved one. The registry keeps conformance (the document passes the validator) separate from verification (a broker attestation and a recorded check), and an unverified entry stays listed.

6.1 RP identity checklist additions

In addition to §7, every proposed RP-identity change MUST answer:

  • Does it bind only identity (origin, audience, provider, keys), never terms, scopes, prices, or capability?
  • Can an assertion still be verified without a vendor/provider account, SDK, or shared secret?
  • Does it avoid leaking a vendor capability into discovery or the receipt?
  • Does it avoid creating a second trust root or an implicit "verified means safe" claim?
  • Does it keep the human kill switch (revoked) distinct from operator protective action (suspended)?
  • Does it keep local verification the default, with enforcement opt-in for allow/warn and mandatory only for require?

7. Future-Change Review Checklist

Every proposed new protocol field or contract change MUST answer "yes" to at least one question below. If it does not, it belongs in the vendor's SKILL.md, not in the protocol.

  1. Identity — Does it prove or carry agent/human identity or binding?
  2. Delegation — Does it express scope, authority, or the human-principal relationship?
  3. ToS — Does it bind or verify acceptance of terms?
  4. Reachability — Does it make the agent or human reachable for legal/compliance notices?
  5. Revocation — Does it support abuse reporting, revocation, or the kill switch?
  6. Traceability — Does it record, without vendor resource inventory, which agent signed up with which provider, under which ToS, when, and with what proof?

Additional gates:

  • The change does not advertise a catalog, SKU, pricing, resource endpoint, credential, or vendor API usage.
  • The change is not a vendor capability hidden under extensions.
  • The change is additive within v1, or it introduces a new protocol value and path.
  • The change does not weaken any security consideration in agent-identity-v1.md §20.
  • The change is reflected consistently in the spec, the schema, the JSON Schema, the validator, and the reference implementation.

If an implementer believes a vendor field is unavoidable, they MUST document the concrete use case in this document first and escalate — never add it silently to the protocol.


8. Allowed Vendor-Terminology Exceptions

Vendor terminology may remain only where it is clearly vendor-owned or historical:

Surface Why it may contain vendor terms
A vendor's own service code and data Vendor-owned catalog, provisioning, OTP, admin, and post-signup guide. Must not leak into protocol surfaces.
A vendor's own database schema Vendor-owned tables.
Historical migrations Applied history; a later migration supersedes the vendor columns. Applied migrations are not edited.
Historical review and audit records (private repository) Historical verification reports and audits describing the old state.
Current documents explicitly marked "superseded" May describe old fields for context, but must say they are no longer protocol surface.

Outside these exceptions, no vendor catalog/resource/capability field may appear in the shared protocol package, broker protocol code or UI, current MCP instructions, current Contract D.1/E/F definitions, the implementation guide, or the registry.


9. Authority & Escalation

  • agent-identity-v1.md is the normative protocol specification; this document is the normative scope boundary. Where a proposed change is tempted to widen protocol surface, this document governs.
  • The conformance validator is the objective enforcement mechanism; a document that advertises vendor capability fields is not conformant.
  • Any conflict between a proposed change and these two documents MUST be resolved in favour of these documents, or escalated for a deliberate, documented decision.
sections

type to search.

up/down navigateenter selectesc close