KYC checks a customer. Know Your Agent (KYA) checks which AI agent is acting, who answers for it and what it may do. Visa, Mastercard, Akamai and Experian now offer ways to run that check, and an Ethereum agent registry went live on mainnet in January 2026. This explainer compares the on-chain and payment-network routes, and shows where both stop.
A bank can ask a person for a passport. An AI agent has no passport and no legal personhood, so it cannot sign a contract or carry liability. As agents start to buy and pay on their own in the agentic economy, KYA schemes share one basic job: tying the software to a person or company that can be held responsible.
What is Know Your Agent?
KYA has no single agreed standard yet, but early frameworks tend to circle the same five questions:
- Which agent is this, and which version?
- Who owns or sponsors it?
- What spending limits and permissions govern it?
- What tools and data can it reach?
- Can its past decisions be proven afterwards?
The answers matter more as the stakes rise. A read-only research assistant needs a lighter check than an agent authorised to move treasury funds or sign purchase orders. That gradation is a working rule of thumb rather than an agreed industry scale.
How is KYA different from KYC?
| KYC (Know Your Customer) | KYA (Know Your Agent) | |
|---|---|---|
| Verifies | A person or company | An agent, plus the person or company behind it |
| Proof of identity | Passport, licence, company registry | Cryptographic keys, signed credentials, an agent registry entry |
| Trust signal | Credit history, background checks | Behavioural reputation, attestations from third parties |
| Authorisation | PINs, one-time codes, signatures | Session keys, spending caps, scoped payment tokens |
| Enforcement | Audits, frozen accounts | Revoked credentials, contract-level limits, optional staking penalties |
| Liability | The account holder | A person or company accountable for the agent’s actions |
How KYA works on-chain: ERC-8004
ERC-8004 gives each agent an on-chain identity in the form of an ERC-721 token. Its specification defines three registries. The identity registry records the agent, the reputation registry collects feedback, and the validation registry logs independent checks of its work. The authors include Marco De Rossi of MetaMask and Erik Reppel.
The proposal is still formally a draft. That has not kept it off mainnet: its reference registries went live on Ethereum on 29 January 2026, followed by deployments on Base, Polygon, Arbitrum and Mantle. Industry trackers put registrations in the tens of thousands, though no primary on-chain count has been published.
ERC-8004 does not make developers post a bond that is forfeited when an agent misbehaves. The standard leaves staking and slashing outside its scope, so that specialised networks can add them on top. Some do.
Wallets add a second layer of control. Coinbase’s Agentic Wallets restrict an agent with session keys that reach only approved contracts, with spending budgets, and with keys held in secure hardware away from the model. They cap the damage. They cannot vouch for each payment the agent chooses to approve.
How KYA works off-chain: the payment networks
Card networks and web-infrastructure firms have built their own versions, mostly off-chain:
- Visa’s Trusted Agent Protocol, launched in October 2025, lets a merchant verify a legitimate shopping agent through signed web requests.
- Mastercard Agent Pay issues agentic tokens that bind a card credential to one agent, a merchant scope and a spending limit. Agent Pay for Machines extended it in June 2026.
- Skyfire’s KYAPay issues signed tokens stating who the agent is, who authorised it and what it can pay for. Experian now uses it as the identity layer for its own agent trust framework.
- Cloudflare’s signed agents programme lets websites recognise AI agents cryptographically instead of guessing from traffic patterns.
In June 2026, Akamai launched an Agentic Security Framework with Visa, Experian and Skyfire as partners. Rubail Birwadker of Visa put the case plainly: “Without trusted identity and explicit permissioning, AI agents cannot participate in commerce at scale.” Skyfire chief executive Amir Sarhangi made the same point: “AI agents can’t participate in the economy without trusted identity and transaction ability.”
Catena Labs, founded by Circle co-founder Sean Neville, raised $30 million in May 2026 to build verified identities, accounts and payment services for agents. The US Office of the Comptroller of the Currency has accepted its application for a national trust bank charter for review. It is the accountable-sponsor model, applied for as a bank.
Where identity checks stop
Security researchers have documented prompt-injection attacks that trick AI agents into making crypto payments. The instructions hide in web pages the agent reads, and the agent obeys them.
This is the limit of identity. Secure hardware keeps a key away from the model, but a manipulated agent can still approve a payment inside its permitted scope. A perfectly identified agent can also be a perfectly fooled one. Spending caps and narrow permissions contain the loss, but they cannot tell whether the latest instruction was honest.
What regulation actually requires
Much of the coverage earlier this year said the EU AI Act’s high-risk rules would bite in August 2026. They did not. The EU’s Digital Omnibus package, in force since 27 July 2026, pushed high-risk obligations back to December 2027 for standalone systems and August 2028 for AI built into regulated products. What did arrive in August 2026 were transparency duties and the AI Office’s enforcement powers over general-purpose model providers. The penalties did not move. Fines under Article 99 still reach €35 million or 7% of global turnover for prohibited practices, and up to €15 million or 3% for breaches of most other obligations.
Payments law runs on its own track. The EU’s PSD3 and Payment Services Regulation are expected around 2027, with no firm start date, and could increase providers’ liability for unauthorised payments. In the US, GENIUS Act stablecoin rules are still in rulemaking, while issuers run identity checks on the person or company funding an agent’s wallet. The old KYC still applies; it simply lands on the human.
No rule names KYA yet. As these regimes take shape, firms will need to show who authorised each agent transaction and how its spending was controlled. That is KYA in all but name.
Frequently asked questions
Is KYA a legal requirement? Not by name. Payment providers must still identify an accountable customer, and emerging payments and AI rules make documented agent permissions more valuable.
What is ERC-8004? A draft Ethereum standard that gives each AI agent an ERC-721 token, backed by registries for identity, reputation and third-party validation. Its registries went live on mainnet in January 2026.
Did the EU AI Act high-risk rules start in August 2026? No. The Digital Omnibus delayed them to December 2027 for standalone systems and August 2028 for AI in regulated products. Transparency duties and AI Office powers over general-purpose AI providers took effect instead.
Can KYA stop an agent being hacked? Not on its own. Spending caps, scoped permissions and key isolation limit exposure, but a manipulated agent can still misuse the permissions it has.
Related on Blockleaders: x402 vs AP2 vs ACP vs MPP: AI agent payment protocols compared and more from our AI section.
