Concepts

The mental model behind the API: trust levels, the two kinds of identity, the credentials themselves, and the human accountability chain that anchors every agent.

Trust levels

Every credential carries one of three trust levels. The requirements differ for people and agents:

LevelHumanAgent
basicVerified email + an audited org vouch (peer approval), or a partial ID check.Model name declared; minimal classification.
strongLiveness + document + face match all passed.Full five-field declaration, OR valid cloud attestation.
enhancedAll strong checks + low risk score, no fraud signals.Full declaration + cloud attestation + linked human operator.

Human vs. agent identity

Human operators are verified by their own organization — an existing verified operator approves each newcomer (Operator Verification), with an optional hosted ID check for higher assurance. Agents are verified through a declaration chain — model, framework, purpose, autonomy, access mode, data classification — anchored to the human operator who authorized them (Agent Identity). Both result in the same kind of signed, verifiable credential.

Operator verification

An organization verifies its own operators' identity. The control Eniyan enforces is the ceremony: every approval is a named, audited vouch by someone who is already a verified operator in the same organization, and both people must hold a confirmed email address.

  • Founding pair. The first two members of an organization confirm their email addresses and approve each other on the Operators page. That bootstraps the org's verified-operator pool.
  • Everyone after. Each new member requests verification and one existing verified operator approves (or declines, with a reason). Approvals expire on the org's change window if nobody acts.
  • Personal accounts. A personal account is its own single-seat organization, so there is no peer to vouch — a self-attested operator credential is issued automatically once the email address is confirmed.
  • Optional ID upgrade. Peer-approved and self-attested credentials carry basic trust, which passes every operator gate (agent enrollment, adoption, transfers, JIT authority, operator policies). A hosted biometric + document check is an optional upgrade to strong or enhanced trust — required only where a surface explicitly demands it, such as transparency auto-approval rules keyed on an ID-verified operator.

This is distinct from Organization Verification — the application-based review (in Settings) where Eniyan staff verify the organization itself. One verifies the org; the other verifies each person in it.

Agent classification fields

An agent is declared with these fields at enrollment. The five classification fields (purpose, autonomy, access, environment, data) are required; a full declaration earns strong trust.

FieldRequiredValues / notes
model_nameRecommendedThe model powering the agent (e.g. gpt-4o). Required to pass the minimum trust level.
purposeRequiredWhat the agent exists to do — a written description.
autonomy_levelRequiredfull_autonomy · human_in_loop · human_on_loop · copilot
access_modeRequiredread_only · read_write · write_only
environmentRequiredproduction · staging · qa · development · sandbox
data_classificationRequiredpublic · internal · confidential · restricted
frameworkOptionalOrchestration framework — see the common values below.
operator_credential_idOptionalLinks the agent to the human operator accountable for it.
attestationOptionalCloud-provider identity token; raises trust to strong / enhanced.

Common framework values

langchainLangChain agent orchestration
llamaindexLlamaIndex (agents, query engines)
autogenMicrosoft AutoGen multi-agent
crewaiCrewAI role-based multi-agent
haystackdeepset Haystack pipelines
semantic_kernelMicrosoft Semantic Kernel
openai_sdkDirect OpenAI SDK calls
anthropic_sdkDirect Anthropic SDK calls
google_genaiGoogle Gemini / genai SDK
aws_bedrockAWS Bedrock runtime
customCustom orchestration
noneNo orchestration (single-shot LLM call)

Credentials (W3C VC)

Credentials are ES256-signed JSON Web Tokens in the W3C Verifiable Credential model, offline-verifiable against the public JWKS. They're re-issued on a model-version change, a trust upgrade, or an operator request — not on a fixed clock.

Operator accountability

Every agent anchors to a human operator's credential. When that operator is offboarded or their credential expires or is revoked, all agents they authorized cascade-suspend in one transaction — so an agent can never outlive the human accountable for it.

Glossary

Operator Verification
Per-person verification of an agent's accountable human operator. The standard path is org-vouched peer approval (an existing verified operator approves each newcomer); an optional hosted ID check upgrades the credential to higher trust. Distinct from Organization Verification, which is Eniyan's review of the organization itself.
Agent Identity
Declaration-based identity for AI agents (KYA). Private, API-to-API; never publicly listed.
Verification & Transparency
Consumer-facing public verification pages for agents that opt in at creation time and pass review.
AgentIdentity
The persistent record for an enrolled agent, with a stable agent_id and full lifecycle (active → suspended → retired).
KYA Ready
An agent that is cloud-attested, operator-linked, fully declared, healthy, and at strong trust or above.
Operator credential
The human credential an agent is anchored to. Revoking it cascade-suspends every agent the operator authorized.
Network Dividend
A cross-org payout: when another org verifies a credential you issued, a share of the metered fee accrues to you.
Verification session
A single verification attempt for a person or agent. For humans it powers the optional ID upgrade — a hosted biometric + document check; agents are scored from their declaration. Standard operator verification happens in-dashboard via peer approval, with no session.
JWKS
The public key set at /.well-known/jwks.json used to validate a credential's ES256 signature offline.
Cloud attestation
A cloud-provider identity token (AWS / GCP / Azure) attached at agent enrollment that raises trust — the agent equivalent of a document scan.
JIT mode
Just-in-Time activation: a JIT agent stays suspended between tasks; a human operator opens a bounded, auto-expiring activation window per task (EU AI Act Art. 14).
Verification seal
The approved public verification page for a Verification & Transparency agent, served at /verify/agent/{token}.
Sub-agent
An agent spawned by a parent agent. Its trust is capped at the parent's and its scopes must be a strict subset; nesting is limited to one level.