Last week I tested two personal AI agents, Grok Bot and Meta Muse. This week, part two: what happens when one of them turns up at your website with a shopping list.

On October 6, Sierra and Meta announced the Personal Agent Protocol, nicknamed "Poppy." It's an open standard for how a customer's AI agent introduces itself to a business, gets permission, and gets things done. Three days later they published the draft and added 35 more design partners, including Target, Gap Inc., Nordstrom, Klaviyo, Visa, Mastercard, PayPal and OpenAI. The shelf, the card networks, the email platform and one of the biggest AI labs are all busy inventing a new game, and you, the people actually selling through them, aren't aware. Luckily, you read The Context Window!

Hashi's take: Your website was built for humans with thumbs who care about beautiful images and clever copy. Its next wave of customers, personal AI agents, won't give either a second glance. What they'll care about: clean structure, accurate pricing and inventory, and clearly written rules of engagement.

If your ecommerce lead still thinks "AI traffic" means a bot scraping your prices, forward this their way. Share your personal link and when they join, you earn rewards: an AI Toolkit Cheat Sheet, an exclusive Deep Dive report, and a 1-on-1 AI strategy chat.

Today, when your agent buys something, it usually logs in as you and clicks around pages designed for people. The business has no idea a bot is driving. It's a bit like sending your nephew to the bank with your debit card and PIN and hoping for the best. Good lad. Wouldn't recommend it.

Poppy tries to fix that with five pieces, which I'll spare you in full:

Sierra's own example is the one every brand should read twice: a retailer lets the agent start a return but not change the payment details on file. That single line is the whole idea.

The demo was Meta's Muse (yes, last week's test subject) going to Rocket's website, finding Rocket's agent, asking permission, and getting its human a mortgage pre-approval in moments. He signed in once to share his credit and income details, and the two agents worked through the price, down payment and interest rate between themselves.

If you sell through Walmart or Target, they're now helping write the rules for how agents shop. Their agent policies will shape your digital shelf whether you join anything or not.

And the agent won't care about your packaging, your end-cap or your clever tagline. It'll care about clean product data, clear policies, and whether your site lets it finish the job.

For a challenger brand, that's a mixed bag. The risk is that the agent defaults to the cheapest or best-known name on the shelf. The upside is that a bot has no idea who paid for eye-level placement. It just rewards whoever answered its questions properly. I've said before that AI is mass-producing expert customers. Now some of them don't even have a pulse.

This is where I need you to pay close attention: writing your poppy.json is not an IT ticket to be closed by Friday. You're deciding what software may do with your customers' money and accounts. That's policy. Somebody senior should own it, and it should go through rigorous scrutiny, ideally from an AI board that reviews it.

A sensible starting list might look like this:

To be clear, nobody needs to publish a poppy.json this month. It's a 0.1 draft, closer to an llms.txt than a deadline. The guest list is the part worth writing now, because you'll need it whichever standard wins. But you should start discussing it internally.

My rule for agents hasn't changed: always make it ask before sending anything into the real world. Poppy gives brands a way to enforce that from their side of the counter, which is more than most of us have today. And yes, I'll keep banging the drum. We spend a fortune making agents clever and pennies deciding what they're allowed to do.

Let me set a little reality check before you go running and rebuilding your checkout around this.

It's a draft. Version 0.1. The protocol site says any part can change before 1.0, including in ways that aren't backward compatible.

It doesn't do payments yet. Payments are listed as an open topic on the path to 1.0, along with push notifications ("your order shipped") and attachments like receipts and return labels. For now, checkout still happens outside the protocol. That's a rather large hole in a shopping story. Rather like a shop with no tills.

Nobody big is live yet. As of this week, people tracking the rollout were still waiting for the first launch or design partner to publish a poppy.json. Sierra says workshops and a reference implementation arrive "over the next month."

There's more than one standard. Visa already runs a Trusted Agent Protocol, and Stripe and Shopify are in both camps. Sierra calls Poppy complementary to shopping-focused standards like the Universal Commerce Protocol. And with Visa now a Poppy design partner too, "pick a side" may turn out to be the wrong question. Everyone, it seems, is backing every horse.

It's too early to pick a side. The best thing you can do if you're a CMO or other decision maker is keep an eye on the landscape and start the conversations around a what-if scenario.

Hit reply: what's the one thing you'd never let an AI agent do on your site?

In July I argued that personal language models are the future: an AI that knows you, works for you, and goes where you send it. Poppy is the first serious attempt at the doorway it walks through. Mind the step.

Protocols like this are going to speed up adoption of, and access to, personal AI.

But with personal AI adoption, security risks are coming, and they will be in places we've not thought about before.

Keep reading!

Grok Bot helped with the research and threw me some writing ideas. The testing and opinions are mine.

- Hashi