Know Your Agent proves the enrollment. It can't prove the moment.
Since 2025, card networks, identity vendors and agent-security companies have converged on the same idea: as AI agents start to transact, you need to know your agent the way you already know your customer. The frameworks are real and they are well built. Card networks now bind an agent to an enrolled account holder, with scoped permissions the person set in advance.
All of that is established at enrollment. It tells you a person once consented to this agent operating within these limits. What it cannot tell you is whether a live human was actually present for the thing that just happened — three months later, on a device that changed hands, under a session someone else may now be driving.
5arz issues the other measurement: a signed credential that a real, ID-verified, live human was present for a specific verified session, checkable by anyone against a public key. The two layers answer different questions and most serious flows will want both.
Consent given once is not the same fact as a person being there now.
Enrollment-time binding decays in exactly the ways credentials always decay: accounts get resold, shared, inherited, or quietly handed to a script. Every agent framework acknowledges this and none of them solve it, because solving it requires measuring a human at the moment of the event rather than at the moment of setup.
The two measurements, side by side
This is not a competitor comparison. These are two different facts, established at two different times, by two different mechanisms.
| KYA — Know Your Agent | Human attestation — 5arz | |
|---|---|---|
| Fact established | This agent is enrolled to this account holder, within these limits | A real, ID-verified, live human was present for this session |
| Established when | At enrollment, then carried forward | At the moment of the verified session |
| Decays if the account is shared or resold | Yes — the enrollment travels with the account | No — a fresh credential is issued per verified session |
| Who can check it | Varies by scheme — often against the issuing network's directory | Anyone — public JWKS, local check, no account with us |
| Privacy posture | Identity known to the enrolling institution | Pseudonymous subject hash in the credential |
| What it is good for | Scope, spend limits, publisher trust, revocation of the agent | Evidence a person was there for this verified session |
5arz does not replace agent identity or KYC, and does not try to. It measures the one thing those layers assume and then leave to you.
Where enrollment-time proof runs out
The account changed hands
The agent is still validly enrolled. The person who enrolled it is not the person operating it. Nothing in the enrollment record can surface that, because nothing re-measures.
Applications and signups at scale
Agents make applying nearly free. Once every submission is legitimate software with a valid enrollment, volume stops carrying information — unless each one arrives with fresh evidence that a distinct real person was present.
Approvals inside multi-agent systems
A human approves a plan; sub-agents execute a dozen concrete actions nobody reviewed. It audits as oversight. Measuring at the step with the irreversible side effect is what closes it.
Human review that quietly isn't
When an agent can pay for "the human step" to be done, your review gate can become another model's output. A per-session human attestation is what makes that difference visible.
How it works
On-device liveness and attention. 5arz stores no biometric templates — only a signed attestation leaves the device.
A short-lived ES256 credential is issued for that verified session, carrying a pseudonymous subject hash and the task type.
Your endpoint, or the counterparty's, receives the credential alongside whatever the agent is doing.
Fetch our public keys once, cache them, then check signatures locally. No SDK, no account, no per-check call to us.
Credentials are short-lived by design and carry an expiry. That is deliberate: a signature check is offline by nature and cannot consult a revocation list, so freshness is enforced by the expiry rather than by a callback. If you need a hard real-time kill switch, use your own agent-side controls or your KYA provider's — that is what enrollment-time frameworks are genuinely good at.
Verify it without trusting us
The point of a signed credential is that it does not require you to believe the issuer. Fetch the keys, cache them, and every later check is local arithmetic. Re-fetch when keys rotate.
The subject is a hash, not a name. It gives you a stable reference and continuity without handing you an identity you then have to protect. Where a lawful process requires resolution, that runs through 5arz rather than through the credential.
A credential is a bearer artefact: record jti on your side and reject reuse, and treat exp as the freshness bound. That single check is what keeps a valid credential from being replayed across sessions.
There is also a live MCP server and an OpenAPI 3.1 description, so an agent can discover and register without a human wiring it up.
Where AI transparency law is pointing
EU AI Act Article 50 transparency obligations became applicable on 2 August 2026. They apply to providers and deployers of specific system classes — chatbots, synthetic-media generators, emotion-recognition and biometric-categorisation systems, deepfake tools — not to AI generally. The marking obligation for generative systems already on the market runs on a transition to 2 December 2026.
California's AI Transparency Act (SB 942, as amended by AB 853) became operative on the same date. It binds covered providers of generative AI services above roughly one million monthly users in California, and covers image, video and audio — not text.
Both mark what a machine produced. Neither requires anyone to produce evidence of what a person did — and in a dispute, that is the artefact that carries weight. Marking machine output tells you what to discount; a human attestation tells you what you can rely on. 5arz does not sell an Article 50 or SB 942 compliance product and this page makes no such claim.
General information, not legal advice. Confirm your own obligations with counsel.
FAQ
What is Know Your Agent (KYA)?
The practice of identifying and vetting an AI agent acting on a user's behalf — what software it is, who published it, which account holder enrolled it, and what it may do. Card networks and identity vendors have built it out through 2025 and 2026 as the agent-era counterpart to Know Your Customer.
How is this different from bot detection?
Bot detection is a negative, probabilistic judgment: it estimates whether traffic looks automated, and every signal it uses becomes a training objective for the next model. A human attestation is a positive, cryptographic claim: a real, ID-verified, live human was present for this session, and the signature either validates or it does not.
Do I have to choose between KYA and a human attestation?
No. Agent identity establishes that the software is what it claims and that an enrolled account holder authorized it. A human attestation establishes that a live person was actually present for a given session. They fail in different directions, which is why they compose well.
What if there is genuinely no human — a fully autonomous agent?
Then no attestation exists, and its absence is the useful signal. The failure mode to avoid is a system that cannot distinguish between a human having been present and nobody having been asked.
Is biometric data stored?
5arz stores no biometric templates. Liveness runs on the person's own device and only a signed attestation leaves it. Document-based identity verification, where a flow requires it, is performed by a regulated third-party identity provider under that provider's own retention policy.
How do I verify a credential?
It is a standard ES256 JWT. Fetch the public keys once from api.5arz.com/.well-known/jwks.json, cache them, and verify with any JWT library. After the key fetch the check is local and typically takes under a millisecond. Re-fetch on key rotation.
Know the agent. Then measure the human.
Free to start, two fields, no card. Or let your agent register itself through MCP.