AI agents do not need a wallet. They need governance.

The instinct, when you first think about how AI agents should pay for things, is to give them a wallet. It is wrong. An agent with its own wallet is an agent with no seatbelt. What agents actually need is scoped, policy-bound access to a wallet someone else custodies.

Every week we talk to a team building autonomous agents and the conversation starts in the same place: "we just need to top up a wallet for the agent." Then they describe what they actually want, and it is never just a wallet. It is monthly budgets, per-call caps, an allowlist of merchants, a way to pause an agent without rotating keys, a receipt the finance team can reconcile, and a human in the loop for anything above a threshold.

All of that lives above the wallet, not inside it. And that is the layer most agent stacks are missing today.

What "give the agent a wallet" actually breaks

• No spending policy. A wallet has a balance, not a budget. There is nothing in a private key that knows the agent is only allowed to spend 500 USDC this month, or no more than 25 USDC on any single call.
• No merchant trust model. The wallet will sign anything the agent asks it to sign. A spoofed x402 merchant or a prompt-injected URL is a direct line to your funds.
• No human in the loop. Approval flows for large or unusual payments require something stateful sitting between the agent and the wallet. A wallet alone cannot pause and wait for a human.
• No clean revocation. If something goes wrong, your only blunt instrument is rotating the wallet's keys, which breaks every agent using it.
• No audit trail. A list of signed transactions is not the same as a list of receipts tied to the tasks that produced them.

Why a governance layer, separate from the wallet

Wallet providers like Coinbase are very good at custody, signing, and on-chain settlement. They are not in the business of expressing your company's spending policies. Mixing those two responsibilities in one place gives you a worse wallet and a worse policy engine.

A clean split looks like this: funds stay with the wallet provider, and a governance layer sits in front of it, holding the rules. Every time an agent wants to pay, the governance layer evaluates the request against the rules, asks the wallet to sign only when it should, and writes a receipt back. The wallet provider never has to know about budgets. The policy engine never has to hold a private key.

What this unlocks

• You can grant a new agent access in seconds without minting a new wallet.
• You can pause or kill an agent's spend without touching the underlying funds.
• You can change policy across a fleet of agents in one place.
• You can give finance an audit log of receipts, not raw on-chain transactions.
• You can swap wallet providers later without rewriting your agents.

That is the layer Vayr Pay is building. Non-custodial by design, x402-compliant, and opinionated about one thing: agents should never be the ones holding the funds they spend.

• Cloudflare will let anyone charge agents. Someone still has to let agents pay.
  Cloudflare's Monetization Gateway turns any API, page, dataset, or MCP tool into a paid resource over x402. That solves the merchant side. The buyer side — how an AI agent actually decides to pay, under whose rules — is the other half of the problem.
• MCP meets x402: how tool-calling agents actually pay for tools
  MCP gave agents a standard way to call tools. x402 gives those tools a standard way to charge. Here is how the two fit together, and what a paid MCP tool call actually looks like end to end.