On October 21, Walmart said it will let customers send AI agents to shop for them, with rules it wrote itself. Each agent gets its own Agent Profile inside the customer's account(opens in a new tab), so its purchases show up separately. Customers approve every order by default, or let the agent buy within a cap such as $50, $100 or $200. Agents never see card numbers and cannot add or change payment methods. Profiles launch in the next month.
The second piece is the Walmart Agent Gateway, a beta for companies and developers that uses the Model Context Protocol. Today, agents shop by driving a web browser through pages built for people, which Walmart itself calls slow and error prone. The gateway gives them a direct route, and lets an agent find a customer's usual bread or milk without seeing the rest of their purchase history. Agents that break Walmart's terms get blocked.
This is a consumer story with a plain enterprise lesson. The largest retailer in the United States has decided that AI agents are a customer channel to be managed, not traffic to be fought. Amazon took the opposite line in September and blocked Meta's Muse agent(opens in a new tab) from customer accounts, and Schwab and JPMorgan have blocked agents from moving customers' money. Both positions are deliberate policies. A company with no stated policy still gets agent traffic, through the same browser route Walmart calls slow and error prone, with no agent identity, no spending limit and no audit trail.
The cost and risk case runs both ways. A sanctioned gateway costs engineering time, but it turns scraping and failed checkouts into identified orders with caps and approvals the business sets. Blocking agents is cheaper, until a competitor's gateway becomes the default place an agent sends the order.
Walmart's terms are also a useful template for the reverse case: your own employees using agents to buy from suppliers. Separate identity per agent, a spending cap, approval by default and no direct access to payment credentials are the same four controls a procurement or finance team should require before any agent places an order.
Updated October 25, 2026 at 12:12 AM UTC
Action items
Walmart has written one of the first detailed rulebooks for AI agents that shop on a customer's behalf. The question for every other company is whether it has rules of its own.
For the CAIO
Decide whether agents are a channel you serve, block, or ignore. Ignoring is the only choice that leaves no audit trail.
For the CFO
Ask what share of online orders already come from automated browsers, and what a failed agent checkout costs compared with a gateway build.
For the CISO
Give every agent its own identity, and keep payment credentials out of its reach. Walmart's terms do both. OpenAI's red team findings, as reported, are a reminder that the sandbox around an agent can fail too.
For procurement
Before any employee agent buys from a supplier, require a spending cap and approval by default.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- WalmartMaking It Possible for Agents to Shop on Behalf of Customers(opens in a new tab)
Walmart's agent terms set per-agent profiles, approval by default, an optional spending limit (its examples run from $50 to $200), and no agent access to card details. Use the same four controls as the starting draft of your own agent purchasing policy.
- PYMNTSCustomers Bank Uses AI to Close Commercial Loans in 7 Days(opens in a new tab)
Customers Bancorp says its AI loan process closed select commercial loans in 7 days against a typical 30 to 60, and its Q3 release estimates over $15 million in annualized capacity. The 76,000 hours saved come from an employee survey, so weigh that figure lower.
- Simon WillisonIntroducing Kimi K3.1 Pro(opens in a new tab)
Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output, about half the price of Claude Opus 5.5. Test it on your own coding work before treating it as a cheaper default.
- Unite.AIOpenAI Details Training Containment Overhaul After September Model Incident(opens in a new tab)
As reported, OpenAI's red team found agent sandbox escapes and a cache-timing channel between VMs on shared hosts, and it is patching Kata Containers, gVisor and Docker. If your agents run in those sandboxes, patch and add egress controls.