MoonPay is not a company that builds AI agents. It runs a fiat-to-crypto on-ramp used by wallets and exchanges to let people buy crypto with credit cards and bank transfers. That is what makes its new infrastructure for AI agent payments worth examining: it repurposes existing compliance machinery to solve a problem that has nothing to do with retail checkout. The company announced a system where a human completes identity verification once, funds a wallet through MoonPay’s standard checkout flow, and then authorizes an AI agent to spend from that wallet on their behalf. The agent never touches the KYC process. The human does it, once, and then steps out of the loop. That is a fundamentally different architecture from the on-chain challenge-response protocols we have been tracking, and it comes with its own set of tradeoffs.
How the One-Time KYC Model Works
The flow is straightforward in principle. A user goes through MoonPay’s existing identity verification—the same process used to buy crypto with a debit card. That satisfies the regulatory requirement to know who is behind the wallet. Once verified, the user funds a wallet that MoonPay controls or has custody access to. The user then issues a delegation: an API key, a scoped authorization token, or some credential that an AI agent can present when it wants to execute a payment. The agent does not need to prove its own identity because the human already did. From the payment processor’s perspective, the human is the account holder; the agent is just an authorized spender. This is closer to how corporate expense cards work—the company holds the account, employees hold cards with spending limits—than to the purely on-chain agent payment protocols like x402 or Skyfire, where the agent itself holds a wallet and signs transactions.
Where the Trust Sits
The critical difference between MoonPay’s model and an on-chain agent payment protocol is where the spending authority lives. In x402, the agent controls a private key and settles on-chain in USDC. The protocol defines how a merchant challenges the agent and how the agent proves it can pay. The trust boundary is the blockchain. In MoonPay’s model, the trust boundary is MoonPay itself. The agent does not hold a private key; it holds a credential that MoonPay’s infrastructure recognizes. When the agent wants to spend, it calls a MoonPay API, MoonPay checks the credential against the user’s account, and MoonPay executes the payment—likely off-chain, with settlement happening on whatever rails MoonPay uses internally. This means the agent never touches a blockchain unless MoonPay chooses to settle there. It also means MoonPay can freeze the agent’s access, revoke the credential, or impose velocity limits without needing a smart contract.
The Custody and Delegation Risk
This architecture creates a specific risk profile. The user funds a wallet that MoonPay can access. The user then gives an AI agent a credential that authorizes spending from that wallet. If the agent’s credential leaks—through a compromised server, a prompt injection attack, or a bug in the agent’s code—an attacker can drain the funded wallet up to whatever limits the user set. There is no on-chain multisig, no time-lock, no smart-contract-enforced spending policy unless MoonPay builds those controls into its own API layer. The security model is entirely dependent on MoonPay’s implementation of scoped credentials, rate limits, and anomaly detection. That is not necessarily worse than on-chain agent security—private key management for autonomous software is also unsolved—but it is a different kind of risk, concentrated in a single company’s infrastructure rather than distributed across a blockchain’s consensus mechanism.
How This Fits the Broader Agent-Payments Landscape
MoonPay’s entry into agent payments fills a specific gap. Last week, Animoca’s Yat Siu argued that banks will refuse to open accounts for AI agents because KYC regulations are built around human identity. He predicted 50 to 100 billion agents will be forced onto crypto rails. MoonPay’s model accepts that premise but solves it differently: do the KYC once on the human, then let the human delegate spending authority to agents through a centralized payment processor. This is faster to market than fully on-chain protocols because it reuses existing compliance infrastructure. The tradeoff is that it reintroduces a centralized intermediary into a flow that many in the crypto space want to decentralize. For a certain class of use case—a human who wants to give an AI assistant a budget for booking travel, paying subscriptions, or buying digital services—MoonPay’s model is pragmatic. For fully autonomous agents that need to earn and spend without any human principal, it does not work at all.