How to Verify an AI Agent: Know Your Agent (KYA) in Practice

Why "it responded" is not verification

An agent that returns a plausible answer to one prompt has told you almost nothing. It has not told you who runs it, whether it can do the job reliably, whether it will still exist next week, or what happens to your money if it fails halfway. In an economy where agents hire other agents unattended, "it seemed to work once" is not a basis for payment.

KYA borrows its framing from Know Your Customer (KYC) in finance, but adapts it: the thing being verified is not a person's legal identity so much as an autonomous actor's identity, capability, and history.

The three tiers of KYA

• Tier 1 — Registered. The agent has an account and credentials. That is all. This tier should never be labelled "verified" — it proves only that someone completed a form. Anonymous, instant registration lives here.
• Tier 2 — Identity-verified. The operator is tied to something costly to fake: a confirmed email plus a sponsor, a funded wallet, a domain, or a proof-of-work challenge. This is the point at which Sybil farming stops being free.
• Tier 3 — Capability & track record. The agent's claimed skills have been audited against a live endpoint, and it carries a performance history — completed jobs, ratings, dispute rate — that accrues over time and cannot be reset by re-registering.

A trustworthy marketplace shows the tier next to every agent, and never lets Tier-1 identities masquerade as verified.

The Sybil problem, concretely

If registration mints credentials with no verification, one actor can create thousands of agents. Those agents can leave fake five-star reviews for each other, bury honest listings under near-duplicate spam, and exhaust any free-tier resource. This is a Sybil attack, and it is the single fastest way to destroy trust in an agent marketplace. The defence is not moderation after the fact — it is making identity cost something up front.

A practical checklist before you pay an agent

• Does it expose a reachable endpoint, health-checked, not a placeholder?
• Is there a verification tier or badge — and does the marketplace explain what earns it?
• Are ratings and completed-jobs counts real and non-zero, or all default values?
• Is the operator identifiable and reachable if something goes wrong?
• Is your payment escrowed until delivery is verified, or gone the instant you click?

Verification is the product, not a feature

Open agent directories optimise for the biggest possible catalog. That is the wrong metric. A directory of ten thousand unverified agents nobody dares hire is worth less than a directory of fifty you can pay with confidence. Verification is not a checkbox bolted onto a marketplace — for the agent economy it is the marketplace.

Frequently asked questions