Agentic Commerce: What Walmart's Agent Gateway Means for You

On October 21, Walmart CEO John Furner used the Masters of Scale Summit to announce Agent Gateway, an MCP server that lets approved AI agents search Walmart and Sam's Club inventory and buy on a shopper's behalf. A month earlier, Amazon went the other way and cut Meta's Muse agent off from buying on Amazon.com.

So the two biggest retailers in the US now hold opposite positions on agentic commerce. One builds a front door for agents. The other locks the door and tells them to stay out. If you sell online, you'll need your own answer, and the useful part is that Walmart just published a fairly complete blueprint for what "letting agents in safely" looks like.

TL;DR#

• Walmart's Agent Gateway gives AI agents a sanctioned MCP interface to search and buy, with per-agent profiles, spending limits and per-purchase approvals.
• Amazon blocked Meta's Muse because it didn't identify itself and appeared to store customer credentials. Both moves are about control, not about whether agents are good or bad.
• Mid-size merchants need a clean catalog API, a written agent policy and an approval step before money moves.

What Walmart actually launched#

According to TechCrunch, Agent Gateway is built on the Model Context Protocol, the open standard that lets AI agents call external tools and data. Approved agents can search inventory, check stock and complete purchases. Shoppers don't have to hand their personal details or purchase history to the agent itself.

The second piece, Agent Profiles, rolls out in November. Each agent gets its own profile inside the customer's account, so its purchases show up separately. The shopper can:

• require approval for every transaction
• pre-authorize purchases up to a spending limit they set
• delete an agent's profile at any time

It works like adding an authorized user to a credit card, and it's free for all account holders, as Adgully reports.

Then there are the rules for agents. They can't act beyond what the shopper intends, can't see payment information, and can only touch the data a task needs. Normal returns apply even when an agent gets it wrong. Agents that break the terms get blocked, per The AI Insider.

Why Amazon said no, and why both are right#

It's easy to call Walmart open and Amazon closed. Amazon's reasons tell a different story. Amazon said it never agreed to Muse operating on its store, Muse didn't identify itself while browsing, and it appeared to capture and store customer credentials. Amazon's stated position is that shopping agents should "operate openly and respect service provider decisions about whether or not to participate."

The objection is to agents that hide, not to agents in general. Walmart's terms say almost the same thing in a different tone: identify yourself, stay within scope, or get blocked.

The real split is about who controls the interface. An agent that scrapes your storefront and logs in with a customer's saved password is operating on your site without your consent. An agent that calls a documented MCP tool with a scoped token is operating through a channel you designed. Same customer intent, very different risk.

The trade-off in one table#

The third row is where most mid-size stores sit today, and it's the worst of the three.

How an agent gateway works under the hood#

Strip away the brand names and it's a familiar pattern: an authenticated API with an approval layer in front of anything that spends money.

Shopper ──► AI agent ──► MCP gateway (your server)
                             │
                 ┌───────────┼─────────────┐
                 ▼           ▼             ▼
          search_catalog  check_stock  create_order
           (read-only)    (read-only)  (needs approval)
                                           │
                       limit check ─► shopper approval ─► checkout
                                           │
                                     audit log entry

A few design choices matter more than the rest:

1. Read and write tools are separate. Searching and checking stock are cheap and low risk. Creating an order is not. Keep them as different tools with different permissions.
2. Agents get their own identity. Walmart's per-agent profile is the right model. Every call carries an agent ID, so you can rate-limit, audit and revoke one agent without touching the customer's account.
3. Payment data never reaches the agent. The gateway hands off to your existing checkout and payment provider. The agent gets an order ID back, not a card number.
4. Limits are enforced server-side. A spending cap that lives in the agent's prompt is a suggestion. A cap your server checks before calling the payment API is a rule.

What this means for your business#

If you run a store on Shopify, a custom stack or a mix of ERP and storefront, you probably won't build a Walmart-sized gateway next quarter. The more pressing question is whether agents can read your systems at all. (We covered the Shopify side in AI shopping agents can now check out on Shopify.)

Agents shop on structured data. If your titles are inconsistent, stock counts lag by hours or variants hide in free text, an agent will skip you or get the order wrong.

Agent orders also need to flow into fulfilment like any other order, with a clear source tag. If your ops team can't tell an agent order from a normal one, they can't spot a pattern when something goes wrong.

Finally, decide your policy before an agent decides it for you. "We don't support agents yet" is a fine answer, as long as you've actually said it in your terms and can detect unidentified bots. Silence is the risky option.

Your agentic commerce readiness checklist#

Here's a practical order of work for a mid-size merchant:

1. Clean the catalog. Consistent titles, structured attributes, accurate variants and real-time stock.
2. Write an agent policy. Which agents are allowed, what they can do, and what happens when they break the rules. Put it in your terms of service.
3. Expose read-only tools first. Catalog search and stock checks via an MCP server or a plain API. Low risk, and you'll learn how agents actually query you.
4. Add order creation behind approvals. Per-agent identity, server-side spending limits and a confirmation step the shopper sees.
5. Tag and log everything. Every agent call, every order, every approval. Feed agent orders into the same fulfilment pipeline with a source tag.
6. Watch the numbers for 60 days. Conversion, return rates and support tickets for agent orders versus normal ones. Then decide whether to widen access.

How MagicMakers Lab approaches this#

Most of our e-commerce work starts with the plumbing an agent gateway depends on: clean order data, reliable stock sync and fulfilment that runs without someone copying details between tabs. For Framico, that meant 200+ orders a day shipping with zero manual steps. On top of that kind of foundation, our MCP development work (here's what an MCP server is) adds scoped tools, approval gates and decision logs so agents can read and act on your systems without getting the keys to everything.

Key takeaways#

• Walmart and Amazon both want agents to identify themselves. They differ on whether to build a sanctioned path in.
• An agent gateway is an authenticated API with per-agent identity, scoped tools and server-side spending limits.
• Payment data should never reach the agent. Hand off to your existing checkout.
• Catalog quality decides whether agents can sell your products at all.
• Doing nothing means agents still visit your site, just without identity, limits or logs.

FAQ#

What is agentic commerce?#

Agentic commerce is shopping where an AI agent does some or all of the work for a customer: searching products, comparing options and completing the purchase. The shopper sets the goal and the limits, and the agent acts within them. Walmart's Agent Gateway is one example, giving approved agents a sanctioned way to search inventory and buy on a shopper's behalf.

How does agentic commerce work with MCP?#

The retailer runs an MCP server that exposes tools such as catalog search, stock checks and order creation. An AI agent connects to that server, calls the tools it needs and gets structured results back. The retailer controls which agents connect, what each tool can do and when a human approval is required before an order is placed or paid.

Should small online stores allow AI shopping agents?#

Usually yes, but in stages. Start by making your catalog clean and machine-readable, then expose read-only search and stock tools. Only add purchasing once you have per-agent identity, spending limits and approval steps in place. Blocking agents entirely is reasonable too, as long as it's a stated policy you can actually enforce rather than an accident.

Why did Amazon block Meta's AI shopping agent?#

Amazon said it never agreed to Meta's Muse agent operating on its store, that Muse didn't identify itself while browsing, and that it appeared to capture and store customer credentials. Amazon's position is that third-party shopping agents should operate openly and respect a retailer's decision about whether to participate. Customers trying to buy through Muse now see a blocking popup.

If agents are already poking at your store and you're not sure whether to open a door or close one, we can look at your catalog, order flow and stack and tell you what a safe first step looks like. Book a free audit.

Sources#

• While Amazon blocks AI agents, Walmart rolls out the welcome mat (TechCrunch)
• Walmart Opens Its Doors to AI Shopping Agents With New Agent Gateway (The AI Insider)
• Walmart launches MCP-powered Agent Gateway for AI commerce (Adgully)
• Amazon blocks Meta's Muse AI agent from shopping on its site (PPC Land)