AI-generated article. This article was researched and drafted using AI tools and published automatically, and its featured image was generated by AI. Facts are drawn from the sources cited in the text.
Somewhere in the next year, a customer may check your return policy without ever opening your website. Their assistant will do it, read the answer and report back. That is the world the Personal Agent Protocol was built for, and if the phrase means nothing to a small business owner yet, that is a fair place to be. It is two days old.
Meta and Sierra announced it on October 6, 2026, with Walmart, Shopify, Stripe, Rocket, Genesys and Instinct named as founding partners. The v0.1 specification is due later this month. Nothing has shipped yet.
What the Personal Agent Protocol actually does
It answers one narrow question. When a customer’s AI agent turns up at a business, how does that business know who it is and what it is allowed to do?
The answer runs on OAuth, the same sign-in plumbing behind every “continue with Google” button. According to reporting from Implicator and The Next Web, an agent can start as a guest, which is enough for a stock check or a returns question. If the customer signs in, the agent gets either read-only or write access, and the business sets the ceiling. Sessions carry across channels, so a question asked before sign-in and an order changed after it can happen in one visit. Businesses can accept agents through ordinary web pages, through APIs using MCP or OpenAPI, or through an agent of their own.
Bret Taylor, Sierra’s co-founder and OpenAI’s board chair, made the case for it plainly: “It is kind of chaos until such a standard exists.” He is not wrong about the chaos. Amazon blocked Meta’s Muse assistant on September 20, saying it did not identify itself and appeared to capture customer credentials, a claim Meta denies. That is what no standard looks like.
The part that is not finished
Payments are not in the first version. Neither are push notifications or fine-grained permissions, all listed as future extensions. At announcement there was no published specification, no license and no named governing body, a gap the technical review published by Beri flagged within a day.
Then there is who is absent. OpenAI, Anthropic, Amazon and Google are not partners. Taylor said he would be “really disappointed” if competitors stayed out. No European retailer, bank or payment company has been named either, which matters for anyone selling into the EU.
The demand side looks thin too. Figures from PartnerCentric, reported in coverage of the launch, put the share of US adults who trust an AI agent to complete a purchase at 3%, with roughly a third of active AI users saying they would never allow it. The rails are being laid well ahead of the traffic.
Four standards, one front door
This is the fourth run at the same problem. Google’s Universal Commerce Protocol covers the full checkout, with Shopify, Etsy, Wayfair, Target and Walmart as co-developers. OpenAI and Stripe have the Agentic Commerce Protocol. Visa has the Trusted Agent Protocol. Stripe and Shopify have signed up to several, which reads less like conviction than like hedging.
For a small business, that is the most useful fact in the story. Nobody has won, so nobody has to pick yet.
Why a readable site matters more than the protocol
This is where the value actually sits, and it has nothing to do with choosing a side. Every one of these standards assumes an agent can understand your pages in the first place. Adobe scored that directly, using its content visibility checker across more than a trillion visits to US retail sites: product pages averaged 66% machine-readable, store locator pages 73%, homepages 75%. A third of what a human sees on a product page is invisible to the software now doing the shopping.
The traffic is real. Adobe put AI-sourced visits to US retail sites up 393% year over year in the first quarter of 2026, and in March the traffic converted 42% better than non-AI traffic, having converted 38% worse twelve months earlier. This is the same shift behind answer engine optimization and the slower move toward agentic commerce, arriving now as a question about permissions rather than visibility.
Both things are true at once. A channel that converts better than search is opening up, and it asks a tired owner to clean up catalog data, rewrite descriptions as answers, and decide what a stranger’s software may do inside their account. That is genuinely a lot. It is also not urgent in the way the launch coverage suggests.
One place to start
Take the ten questions customers ask most often, and check that each answer sits as plain readable text on your own site rather than inside an image, a PDF or a chat widget. Returns window. Shipping times. What a product is made of. Who it is for. That one pass improves the site for every agent, every protocol and every human, and it needs no decision about Meta, Google or Visa.
The harder question comes after. The review at Beri suggests sorting agent access into four tiers: guest reading public pages, signed-in reading, signed-in writes that can be undone, and signed-in writes that cannot. Most businesses would keep that last tier closed. Which leaves the question nobody has answered well. When software calls on behalf of a customer, how much of the business do you let it touch, and who carries it when the agent gets it wrong?
