Skip to content

ADR#0036: Agent Self-Certifying Cryptographic Identity ​

Context ​

An agent needs an identity that is provable anywhere, by anyone, without a central authority issuing or vouching for it. ADR#0024 and ADR#0025 define the agent record and its configuration but say nothing about how an agent is cryptographically identified, nor how a consumer proves that an action came from a given agent. Today the genesis AgentProvisioned event carries an opaque agent_id string, display_name, parent, an AgentConfiguration, a revision, and a content_digest. Nothing ties the agent to a verifying key, so the registry holds agents with no cryptographic anchor: an agent_id alone proves nothing.

Two different things need verification, and they are independent:

  • Configuration integrity, "is this the same definition?", answered by content_digest and already present.
  • Actor authenticity, "did this agent produce this?", currently absent.

The prevailing solution for globally provable identity without a central allocator is self-certifying identity: the public key is the name. Nostr (which the Buzz agent platform uses), Bitcoin, and W3C DIDs all take this shape. Two properties fall out of it for free:

  • Global uniqueness: a key from a 256-bit keyspace is unique with overwhelming probability, generated by anyone with no coordinator.
  • Global verifiability: anyone, offline, can verify a signature against the public key; no phone-home to an issuer.

The platform already has the verification machinery. ADR#0017 binds an agent to a cnf.jwk public key and verifies proof-of-possession in trogon-aauth-verify, and ADR#0026 hands the decider a typed CommandPrincipal after authorization runs, so decide never sees keys or signatures. ADR#0023 and ADR#0033 keep private key custody on the security plane. What is missing is the binding at registration: recording the agent's public identity on the genesis event, and requiring the registrant to prove control of the key it claims.

Decision ​

An agent's identity is a self-certifying keypair: the public key is the name.

  1. Self-certifying identifier. The agent's identity is derived from its root public key (a DID / JWK-thumbprint form, aligned with AAuth's agent_jkt; the concrete algorithm, identifier method, and signature format are pinned in ADR#0038), not an opaque allocator-assigned string. Global uniqueness and global verifiability are therefore intrinsic, with no central registry.

  2. One root key in v1. A single root identity key, which is also the signing key. A two-tier hierarchy (an offline root key that only authorizes rotation, plus rotatable operational signing keys) is deferred until key operations justify it; it is a strictly additive layer that does not change the genesis anchor.

  3. Bound at registration by proof-of-possession. Provisioning carries the agent's public key and a signature proving the registrant controls the corresponding private key. That proof is verified on the AAuth plane (ADR#0017) before decide runs (ADR#0026); the genesis event records the public key as the agent's immutable identity anchor. No private key material ever enters a command, event, or AgentConfiguration.

  4. Public identity is a genesis fact; secrets are not. The public key is agent-owned identity and belongs on the agent record. Private keys and custody remain on the security plane (ADR#0023, ADR#0033). This is distinct from the provisioning actor ("who ran the command"), which is envelope metadata, not agent identity.

Deferred layers ​

These do not change the genesis anchor and are added when needed:

  • Entity attestation: binding the key to a real-world entity ("this is Acme's reviewer agent") via verifiable credentials, a web of trust, or reputation accrued in the signed history. The key proves control, not affiliation.
  • Global resolution: publishing the identifier-to-current-key mapping to a replicated substrate (relays, a DID method, a transparency log) so parties outside this system can resolve it. Until then the agent's event stream is the authoritative key record. ADR#0053 settles the narrower external-verifier case with an OIDC issuer and JWKS, which is what non-AAuth parties actually consume; it does not close this layer, because resolving the identifier is the part that still needs a DID method.
  • Key rotation: a later signed event authorized by the current key, with the identifier held stable by the resolution layer above. Pure self-certifying identifiers are immutable, so rotation and resolution arrive together.

Consequences ​

  • An agent is a first-class cryptographic principal: its identity proves itself anywhere, offline, without a central authority, and actions attributed to it are verifiable against the key anchored at genesis.
  • The genesis contract gains the agent's public identity key; agent_id becomes a self-certifying identifier rather than an opaque string. Verification reuses the single AAuth path ADR#0038 fixes rather than a parallel mechanism.
  • The governance half of the picture, decentralized verification under governed authority (admission, revocation, tenancy, attestation), is recorded in ADR#0037; how the anchored identity extends to command and event authorship is recorded in ADR#0039.
  • Global uniqueness and verifiability are obtained by construction; the harder, ongoing work (entity attestation, global resolution, rotation) is isolated in deferred layers that leave the anchor unchanged.
  • Choosing self-certifying identity now forecloses an allocator-assigned opaque identity later without a migration, which is the intended trade.

References ​

Released under the Apache License 2.0.