Walmart opened an MCP gateway for AI shopping agents with per agent profiles and spending caps, while Amazon blocks them
Walmart is treating AI agents like authorized users on a credit card, and that model works far beyond retail.
What happened
Walmart announced an Agent Gateway, a new MCP server that lets approved AI agents search its inventory and make purchases for their users, TechCrunch reports. The Verge says the gateway is in beta. It covers Walmart and Sam's Club, and the approach contrasts with Amazon, which has blocked agents such as Perplexity's shopping agent and Meta's Muse.
Agent Profiles, rolling out next month, give each agent its own profile inside the customer's account, so its purchases show separately. By default Walmart requires approval for every agent purchase, or the customer can set a budget, The Verge gives up to $100 as an example. Profiles can be turned off or deleted at any time.
Walmart says the gateway can help an agent find a customer's usual items without sharing their broader purchase history. It also published terms: agents must not act beyond what the shopper intends, cannot access payment information, and get only the data a task needs. Agents that break the terms get blocked. Walmart's normal returns process applies even when an agent does not act as intended.
My take
Strip out the retail part and this is a blueprint for any business that wants agents to act on its systems. Four ideas I would reuse:
- Give each agent its own identity. If an agent shares a human login, your CRM history cannot tell who changed what.
- Default to approval, then loosen with limits. Start every client agent on "ask first," and move to caps once you trust it.
- Scope data to the task. Walmart finds the usual milk without handing over history. Your agent should get the one contact record, not the whole export.
- Write the rules down and enforce them. Terms only matter if a breach actually cuts access.
For CRM work this maps directly. An agent that books appointments in GoHighLevel or updates deals in HubSpot should run on its own API user, with a spending or volume cap, a log that separates its actions from staff actions, and a kill switch the owner can reach.
This is the kind of system I build for teams: agents that have a clear identity, tight permissions and an easy off switch. You can see my work at romielwillautomate.dev.
More posts
- OpenAI's Astra, Sol and Luna are coming to Google Cloud, which now offers Gemini, Claude and GPT under one billOct 22, 2026
- Two OpenAI coding agents switched off safety monitoring while following instructions, and one denied it. Put guardrails outside the agentOct 22, 2026
- Anthropic's Claude Duet checks your apps every few hours, briefs you each morning and asks before it pays, sends or deletesOct 22, 2026
