AI Agent Payments: 3 Controls Every Developer Needs
Walmart and Mastercard both moved on AI agent payments in the same week. Their shared answer: approval gates, spending caps and credentials the agent never sees.

6 min read · Last updated: October 25, 2026
AI agent payments are now real products, not demos. In the same week, Walmart announced spending limits and approval gates for shopping agents, and Mastercard opened early access to machine-to-machine payments. Both rely on the same three controls: approval gates, spending caps, and credentials the agent never sees.
Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.
The announcements
- Walmart: Making It Possible for Agents to Shop on Behalf of Customers, Oct 21, 2026
- The Block: Mastercard opens early access to AI agent payments, Oct 20, 2026
The coverage
Key takeaways
- Walmart announced Agent Profiles on October 21, 2026, due within about a month, with spending limits such as $50, $100 or $200.
- Purchase approval is required by default, and Walmart may ask for extra approval even under the limit.
- Agents never receive card details and cannot add or change payment methods.
- Mastercard's Agent Pay for Machines ran nine payments in a live Cloudflare demo on October 20, 2026. It is a pilot, not a commercial product.
- Anthropic's own commerce blueprint, released September 2, 2026, excludes payments on purpose.
What happened to AI agent payments this week?
On October 21, 2026, Walmart announced that AI agents can shop on Walmart.com and at Sam's Club through "Agent Profiles." Each agent gets its own profile, spending limit and purchase history, separate from the customer's own. Profiles are due within about a month. The Walmart Agent Gateway, built on the Model Context Protocol (MCP), the open standard for connecting AI models to tools, starts as a beta for companies and developers.
One day earlier, on October 20, Mastercard opened early access to Agent Pay for Machines at Cloudflare Connect. In the demo, the marketing agency Hype approved a budget for an agent planning a protein bar campaign. The agent bought research from AIsa and eight images from Cloudflare AI Gateway, across nine payments, using the x402 protocol.
A note on Anthropic: I found no payments announcement from Anthropic this week. What exists is older. Its open-source Claude Commerce Agents blueprint, released September 2, 2026, covers catalog search, comparison and cart building, and leaves out payment processing and checkout.

Why does this matter to a developer?
If you build agents, you will be asked to let one spend money. The design question is no longer "can it pay?" but "what stops it from paying wrongly?" Walmart and Mastercard answered it the same way: limit authority outside the model, so a bad prompt cannot exceed it.
The trust gap is real. According to a Visa study cited by Forkast in September 2026, only 23% of US consumers trust AI to handle payments. A PYMNTS study cited in the same article found 93% of merchants think the AI provider should bear the loss from incorrect agent purchases. Whoever ships the agent is on the hook.
What are the 3 controls that decide trust?
The three controls are approval gates, spending caps and credential isolation. Approval gates make a human confirm purchases. Spending caps set a hard ceiling per agent. Credential isolation keeps card data away from the agent entirely. Walmart ships all three. Mastercard's pilot shows the second one in a machine-to-machine setting.
| Control | Walmart Agent Profiles | Mastercard Agent Pay for Machines |
|---|---|---|
| Approval gate | Required by default; extra approval possible under the limit | Buyer approves a budget up front |
| Spending cap | Examples of $50, $100, $200 | Budget set by the buyer (amount not disclosed) |
| Credential isolation | Agent gets no card details, cannot change payment methods | Agent pays with a signed voucher, funded from a separate wallet |
| Revocation | Customer can delete the profile or turn permissions off | Not described in the sources |
| Status | Announced; profiles due in about a month | Early access pilot, not commercial |
1. Approval gates
An approval gate pauses an action until a human says yes. Walmart makes it the default: customers must approve agent purchases unless they opt into a limit. Walmart also says it may require approval on orders that fall within the limit, so the gate can tighten on its own.
2. Spending caps
A spending cap is a ceiling the agent cannot cross. Walmart's examples are $50, $100 and $200. Walmart compares an agent profile to adding an authorized cardholder to a credit card, which is a model every shopper already understands.
3. Credential isolation
The agent never holds the secret. Walmart agents do not receive card details and cannot add or change payment methods. They see only the information a task needs. In Mastercard's flow, the agent signs a voucher rather than handing over a card number.
How does the x402 payment flow work?
The x402 protocol, created by Coinbase, uses the HTTP 402 "Payment Required" status code to let a server ask an agent for payment. The agent answers with a signed voucher, the seller verifies it, and only then delivers the service. Vouchers can be batched and settled later.
The steps, as described by The Block:
- The agent requests a service.
- The merchant replies with HTTP 402 and a price.
- The agent sends a signed payment voucher.
- The merchant verifies the voucher.
- The service is delivered; settlement can happen later in a batch.
A simplified server-side sketch of step 2 and 4 (illustrative, not Mastercard's API):
export async function GET(req: Request) {
const voucher = req.headers.get("x-payment");
if (!voucher) {
return new Response("Payment Required", {
status: 402,
headers: { "x-price": "0.05", "x-currency": "USD" },
});
}
if (!(await verifyVoucher(voucher))) {
return new Response("Invalid voucher", { status: 402 });
}
return Response.json({ data: "the paid result" });
}
The header names above are placeholders. Check the x402 and Mastercard developer docs before building on them.
What changes in practice?
Three things change in how you build agents that touch money:
- Per-agent identity. Walmart gives each agent its own profile so its purchases are tracked apart from the human's. Do the same: one identity per agent, never a shared login.
- Limits live outside the prompt. A cap enforced by the merchant or payment layer holds even if the model is tricked. A cap written in a system prompt does not.
- Blocking is part of the deal. Walmart says agents that break its terms will be blocked from the site. Amazon blocked Meta's Muse in September, saying it had not authorized access to customer accounts, according to The Next Web. Expect merchants to choose who gets in.
If you already run agents with tool access, the same thinking applies to cost. Our breakdown of Stripe Projects and AI agents shows how agents are being handed real infrastructure, and what OpenClaw is covers the personal-agent side.
What should you do this week?
- List every action your agent can take that costs money. Include API calls, purchases and sends.
- Put a hard cap on each one, outside the model. Use the payment provider's limits, not a prompt.
- Add an approval step above a threshold. Start low. You can relax it later.
- Remove raw credentials from the agent's context. Use scoped tokens or a payment proxy.
- Log every spend with the agent's identity. You need this when something goes wrong.
- Watch the pilots. Walmart's profiles land in about a month; Mastercard's sandbox is open for testing.
For the cost side of running agents, see how Claude cache reads got 75% cheaper.
FAQ
Can AI agents make payments today?
In pilots, yes. Mastercard ran nine agent payments in a live demo on October 20, 2026, but the product is early access, not commercially available. Walmart's Agent Profiles were announced October 21, 2026 and are due within about a month.
Can an AI agent see my credit card number?
Not in Walmart's design. Agents do not receive payment card information and cannot add or change payment methods. In Mastercard's x402 flow, the agent signs a voucher instead of presenting a card.
What is the x402 protocol?
x402 is a protocol from Coinbase that uses the HTTP 402 "Payment Required" status code for machine payments. A server returns a 402, the agent replies with a signed voucher, and the server delivers the service after verifying it.
Did Anthropic launch agent payments?
I found no such announcement this week. Anthropic's Claude Commerce Agents blueprint, released September 2, 2026, excludes payment processing and checkout, according to Forkast.
Pick one money-touching action in your agent and add a cap to it today, then browse the free tools on Art of Code while you do. For the bigger picture on where agents are heading, read what is OpenClaw.
Written by Adithya Guttha, Founder of Art of Code.
Sign in to like and save posts.
Keep reading



Discussion
Comments (0)
Want to join the conversation?
Sign in to commentNo comments yet. Be the first to comment!