Walmart Let AI Agents Shop. Amazon Didn't.

The split is real and it shipped this week

On October 21, 2026, Walmart announced the launch of an Agent Gateway, a new MCP (Model Context Protocol) server that lets approved AI agents connect to Walmart's systems, search inventory, and complete purchases on a customer's behalf. The same features extend to Sam's Club. You don't need a Walmart+ membership; any account holder can use it.

At nearly the same moment, reporting from both TechCrunch and The Verge confirmed that Amazon is blocking AI agents from making purchases on its platform. Other online businesses, including travel and booking sites, are doing something similar.

So you have the two biggest U.S. retailers pulling in opposite directions on the same question: should an AI agent be a first-class shopper, or a bot to be kept out?

What the Agent Gateway actually is (technically)

Strip the press-release language and the core mechanism is an MCP server. The Model Context Protocol, originally created by Anthropic, is the open standard that lets AI agents talk to external data sources, tools, and workflows. Walmart built a server on top of it so that an agent doesn't have to parse a browser, click through a cart, and fight a CAPTCHA.

The Gateway is in beta. The full consumer-facing rollout, including Agent Profiles (more on that below), is slated for November 2026. So if you're a developer poking at this right now, you're on the early side of the curve.

The practical advantage, as The Verge's Emma Roth framed it, is that using an API-standard protocol sidesteps the hiccups agents hit when navigating a standard browser session. You get structured access instead of screen-scraping chaos.

The guardrails Walmart actually built in

This is the part that matters if you're a developer or founder thinking about building on top of agentic commerce.

  • Agent Profiles. Each agent gets its own profile inside the customer's Walmart account, described by TechCrunch as analogous to adding an authorized user on a credit card. The customer can see which agent made which transaction.
  • Spending limits and transaction approval. By default, Walmart requires the customer to approve every agent-initiated purchase. You can also set a budget cap (The Verge cited an example of up to $100) and let the agent operate within it. TechCrunch notes the customer can also require per-transaction sign-off.
  • Scoped data access. Agents are restricted to only the information needed to complete a specific task. They don't get broad access to a customer's personal data, purchase history, or payment information. Agents are explicitly prohibited from changing payment methods.
  • Compliance or blocking. Agents that violate Walmart's terms for AI agents get blocked from the site.

One important caveat: neither TechCrunch nor The Verge discloses the exact technical mechanism that enforces the data-scoping restrictions. The policy is stated; the implementation details aren't. Treat "agents can't see your purchase history" as a policy commitment you'd want to verify in the API docs, not as an architectural guarantee you've been shown.

Walmart also says its standard returns process applies to agent-initiated purchases, including cases where the agent didn't act as the customer intended. That's a liability posture worth noting.

The vendor question is genuinely unclear

This is where the reporting from the two major outlets directly conflicts, and I'm flagging it because it matters if you're planning an integration.

  • TechCrunch (Sarah Perez) describes Walmart as partnering with Meta's Muse AI agent and offering integrations with ChatGPT, Gemini, and other popular AI apps.
  • The Verge (Emma Roth) reports that Walmart blocked Perplexity's shopping agent and Meta's Muse AI assistant.

You can't reconcile those two from the published articles. One of them has it backwards, or the relationship changed between the time each piece went to press, or "partnering" and "blocked" are describing different scopes of access. Until Walmart publishes a definitive list of supported agents, assume you don't know yet which vendors are cleared.

What both sources do agree on: the Gateway is open to any compliant agent, and non-compliant agents get cut off. The supported-vendor list is a moving target.

Why the Amazon contrast is the more telling signal

Walmart's move is easy to frame as a PR win. Amazon's counter-move is the data point more people are going to underweight.

Amazon is the company with the most sophisticated retail-funnel instrumentation on the planet. If it's actively blocking agent-initiated purchases, the most likely explanations are that it sees agents as a conversion-risk vector (agents don't browse, they don't get upsold, they don't linger), a data-extraction channel (even a scoped agent is a new off-ramp for purchase behavior data), or both. You can infer that Amazon's product team looked at the unit economics of an agent transaction and didn't like the margin profile.

Walmart is the opposite bet: the agent is the customer interface, and the marginal cost of serving a structured API call is lower than the cost of keeping a human in a loop. Whether that bet pays off depends on whether agent-initiated purchases can be steered into the same basket-size and attach-rate patterns that human browsing produces. Nobody knows yet, because the data doesn't exist.

What this means if you're building

A few practical takeaways for developers and technical founders:

  1. MCP is moving from "LLM tool-calling protocol" to "commerce integration standard." Walmart building a first-party MCP server for a two-division retail operation is a stronger adoption signal than any number of developer demos. If you're building agent infrastructure, MCP support is no longer optional table-stakes.
  2. The consent-and-audit pattern is worth copying. Agent Profile + spending cap + per-transaction approval + scoped data access is a clean template for any system where a delegated actor operates on behalf of a principal. Whether it's e-commerce, travel bookings, or B2B procurement, the same four controls probably apply.
  3. Don't build until the vendor list is stable. The TechCrunch/Verge conflict on who's supported means the API surface and access rules may still be shifting. Watch for the November 2026 GA window before committing integration work.
  4. The returns-and-liability language is your due-diligence checklist item. If you're building an agent that will transact on a customer's behalf, have a clear answer for what happens when the agent buys the wrong thing, buys too much, or buys something the customer can't return. Walmart's answer (standard returns apply) is convenient for them. Your answer might need more work.

Open questions nobody has answered yet

  • How does the Gateway technically prevent an agent from accessing personal data or purchase history? Policy is stated; mechanism isn't.
  • What are the exact compliance terms, and what does a violation look like on the wire?
  • What does the agent-facing approval workflow actually feel like in the November rollout? Push notification? In-app prompt? Inline confirmation?
  • Is the Gateway a restricted beta, a partner-only integration, or a broader developer-facing API? The language in both articles hedges.
  • How do returns, refunds, and support tickets flow when the agent, not the human, is the one who made the purchase?

The split between Walmart and Amazon is going to resolve one way or the other over the next 12–18 months. The company that gets the agent-conversion economics right—without gutting its data moat—sets the expectations that every DTC brand, marketplace, and booking site will have to meet. Right now, that company is taking its time, running a beta, and watching the metrics.

And the other one is blocking the bots at the door.

--- Reporting in this article draws on TechCrunch (Sarah Perez, Oct 21, 2026) and The Verge (Emma Roth, Oct 21, 2026). Where the two sources conflict on supported agent vendors, both accounts are noted rather than resolved.