Meta, Walmart and Shopify are writing the rules for AI agents that act for customers
On 6th Oct'26, Sierra announced the Personal Agent Protocol, an open standard for how a customer's own AI agent connects to a business, built with Meta, Walmart, Shopify, Stripe, Genesys, Instinct and Rocket. It runs on OAuth, the same plumbing behind signing in with a Google account: an agent starts as a guest that can check stock or a returns policy, then gets read-only or write access once the customer signs in and grants it, and the business sets what each level is allowed to do. A company can choose whether agents come in through its website, its APIs (including MCP) or an agent of its own. The first draft of the specification, v0.1, is due later in October, and OpenAI, Google and Anthropic are not named as partners.
What it actually means
This is the third time in a week the same question has come up from a different direction. OpenAI's Dots agents can read while you are away but not write, and Meta's Muse agent left a buyer and a seller arguing over who had allowed what. The protocol is the business side's answer: the customer grants access in tiers and the company decides what each tier can do, so a dispute over who authorised an action becomes a record you can check rather than one word against another.
If customers reach you through a website or an app, the work to start now is internal, because the spec is only a promise until the draft lands. Decide which actions you would let an outside agent take on a signed-in customer's account, such as changing a delivery or placing an order, and which still need the customer in person. With none of the big model labs on the partner list, treat this as one candidate standard rather than the settled one.
An agent that acts for your customer needs a permission you can check afterwards.
Source: Sierra · Wednesday 7 October 2026