Walmart will give every shopping agent its own profile. Your agents still share your login.

Next month a Walmart shopper will have tighter control over their shopping agent than most finance teams have over the agent touching their vendor master. The shopper gets a separate profile per agent, a spending cap, per-purchase approval and a delete button. Check what your agent got: probably a shared service account, a stored credential, and every permission that account already carried.

Walmart announced three things on Wednesday 21 October: Agent Profiles, the Walmart Agent Gateway, and written terms for AI agents¹.

The reason Walmart gives for building Agent Profiles is the part worth reading twice. Today, it writes, when a customer uses an AI agent, "it signs in using a customer's account", so all purchases, "whether made by the customer or the agent", appear together, "without any differentiation"¹.

That sentence also describes your audit trail.

/01What the controls actually are

Agent Profiles launch in the next month¹. You authorize each AI agent with a dedicated profile inside your account, so that agent's purchases sit apart from yours. You choose how purchases get approved: require approval for every purchase, or give the agent a spending limit you set, "for example, $50, $100 or $200"¹. Permissions and limits can be turned off at any time, and the profile can be deleted at any time¹.

Walmart's own analogy is adding someone as a cardholder on your credit card¹. The agent gets a way to pay, a cap, and a line of transactions with its name on it.

The written terms read like an approval policy, which is more than most internal agent rules manage. Agents must not take actions beyond what the shopper intends. Agents do not receive payment-card information and cannot add or change a payment method. Agents can access only the information needed for the task and allowed by the customer. Customers must approve agent purchases by default, and Walmart says it may require additional approvals even for purchases inside an agent's spending limit¹. The returns process applies "including if an agent doesn't act as intended"¹.

Then the line that changes your planning: agents that do not follow these terms "will be blocked from Walmart's site"¹.

The Gateway is the carrot. It is an MCP endpoint, starting as a beta for companies and developers, because Walmart says a website built for people to browse and click is "slow and error-prone" for agents, which "may get stuck along the way"¹³. It can also find a customer's usual items without handing the agent their purchase history¹. Amazon went the other way and moved to block shopping agents acting for customers².

/02The same shape shipped twice in two days

On Thursday 22 October, 1Password launched 1Password for Agents. The agent reaches a sign-in form, proves its own identity to 1Password with a unique digital key, and you see which credential it wants and why. You approve on your own device, and 1Password fills the credential into the form instead of sending it to the agent⁴.

Their product chief, Aashish Tripathi, put the old pattern plainly: "The old model where you hand over a credential and hope for the best doesn't work." An agent, he said, should do its job "without getting permanent access to everything behind the account"⁴.

Two companies, two days, one design: identity per agent, approval off the agent's own channel, no standing access. Meanwhile TechCrunch, citing Axios, reports US banks have begun to detect agentic activity inside user accounts, and that most banks cannot tell whether that activity is tied to a user request. Bankers described the situation as "horrifying"⁵.

/03What to change this week

1. /01

   One identity per agent, never a shared service account. If two agents and a nightly script log in as svc-automation, you cannot answer "which agent did this" or revoke one without breaking the others.

2. /02

   Put the cap on the agent, not the team. Walmart caps the profile. Your limit should attach to the agent and the action, in the form "refunds up to $200, existing customers, once per order", so a drifting agent hits a wall its instructions cannot move.

3. /03

   Approve where the agent cannot write. 1Password confirms on the person's device. If the "yes" arrives in a channel the agent can post to, it is not approval.

4. /04

   Make agent actions separable in your own records. Open your ERP change log and try to tell agent writes from human writes. If you cannot, you have the exact problem Walmart just described, on vendor records instead of groceries.

5. /05

   Read your counterparties' agent terms before your agent calls their API. Being cut off is now a normal failure mode. Walmart says non-compliant agents get blocked, so somebody on your side has to own those terms the way they own a contract.

That is the job Gatehouse does. Each agent runs with its own identity and its own limits, anything outside them waits for a named approver, and every action lands on a record you can hand to an auditor or a counterparty.

Pick the agent of yours with the most reach today. Give it a name, a cap, and a line in the log that is only its own.

Want the next one? News with a take, three times a week. Follow by RSS →