Somebody just started writing the rulebook for the AI agents that will knock on your shop's door, and Meta is in the room. On Monday, the customer-service AI company Sierra announced the Personal Agent Protocol, an open standard for how a shopper's personal AI agent talks to a business. Meta helped build it. Walmart, Shopify and Stripe signed on.
That's the lead today, and it sets the tone for the rest: the agents are coming, the plumbing is being laid, and the security story is still catching up. Here's what you need to know.
Meta and Sierra write the rules for shopping agents
The idea is simple. Your customer has an AI assistant. That assistant wants to ask your store something, or do something for them. Today there's no common way for it to prove who it's working for or what it's allowed to touch. The new protocol fixes that with the same sign-in technology (OAuth) that sits behind every "Log in with Google" button.
It works in two levels. As a guest, an agent can ask the easy stuff: is this in stock, what's the returns policy. Once the customer signs in, they decide whether the agent gets read-only access (check my order) or write access (place it, change it). The business decides which features to open up, and can see what the agents are doing, whether they come through the website, an API, or the company's own agent. Sierra calls it a preview for now; a first v0.1 spec is due later this month. The partner list also includes Genesys, Instinct and Rocket.
Why it matters for you: look at what guests can ask. Stock and returns. Those are the questions already flooding your DMs. Soon a chunk of them will come from a bot acting for a person, and the bot will want a clean, correct answer, fast. If your product info lives in your head and in old Story highlights, an agent can't use it. Getting your catalogue, prices and policies written down properly is no longer just tidy housekeeping; it's how you stay findable. We wrote up how to teach an assistant your product catalogue so it stops guessing, and the same discipline applies here.
One honest caveat: this is a standard for big platforms and big retailers first. Small sellers on Instagram won't plug into it next week. But when Meta helps write the rules, they tend to show up inside Meta's apps eventually.
Google's new image model does more for half the price
Also on Monday, Google released Nano Banana 2.1, the next version of its image generation and editing model. The headline is cost: roughly half the price of Nano Banana 2. According to The Decoder, a 1K image drops from about $0.067 to $0.034, and a 4K image from about $0.15 to $0.076.
The upgrades are the ones sellers actually feel. Better text rendering (your price tag or sale banner comes out spelled right more often). Better subject consistency, so the same product or model looks like itself across a series of images. Mask-based editing lets you change one part of a photo and leave the rest alone, and it can take up to 14 reference images at once. It's rolling out across the Gemini app, AI Mode in Search, AI Studio, Flow and Google Ads.
The take: cheaper and more consistent is exactly what a small shop needs for catalogue shots and campaign variations. But don't throw away your camera. The Decoder's own testing found the older Pro model often still looks better, and 2.1 struggles with scale, so a mug can come out the size of a bucket. Use it for backgrounds, banners and variations; keep the real photo of the real product. If you're deciding which AI tools are worth your time, our list of AI tools for Instagram sellers is a good place to start. Note: the previous model is being switched off on October 29, so if a tool you pay for runs on it, expect a change.
Meta's Muse agent had a scary bug before launch
Now the other side of the agent story. 404 Media reported on Sunday that in the weeks before Muse launched, Meta engineers found several security holes in its AI agent. At least one was a "VM escape": a way for an ordinary Muse user to break out of the sandbox the agent runs in and, according to Meta's own documents seen by 404 Media, reach sensitive internal Meta databases.
The bugs were fixed before launch, after what the report describes as a multi-team mad dash that went all the way up to Mark Zuckerberg. Meta now pays up to $300,000 in its bug bounty for a Muse VM escape. That number tells you how seriously they take it.
Why should a seller care about a bug that got patched? Because Muse is the same agent that is learning to shop and check out, and it's not the first time it has made us nervous; last week we covered the time Muse gave away a seller's address. Agents that act on your behalf need access to real data: orders, addresses, sometimes payment. The more an agent can do, the bigger the damage when it breaks. When any AI tool asks for access to your shop, ask what it can see, what it can change, and what happens when something goes wrong.
Mistral ships a giant model you'll be able to own
France's Mistral released a preview of Mistral Large 4, nicknamed "Le Chonk." It's about one trillion parameters, with around 49 billion active per answer, and it handles text and images together. The preview runs through Mistral's own platform, and the weights (the model itself) are due around October 27 under an open license. Mistral trained it in its own European data centres.
Mistral's pitch is blunt: if you own the model, no vendor can switch a capability off on you. For a seller that's not a weekend project; you won't run a trillion-parameter model in your back room. But it matters indirectly. Every serious open model pushes prices down and gives the tools you use more options. More competition at the top usually means cheaper, better AI at the bottom.
Quick takes
- AI is booking real appointments. Flai raised a $27M Series A for its AI dealership software, which TechCrunch reports books about 50,000 sales and service appointments a month across more than 10 of the top 50 US dealer groups. Not a demo. A sales assistant doing the follow-up humans forget.
- Instagram's disappearing photos are under investigation. UK regulator Ofcom opened a probe into whether Meta properly checked the risks of Instagram Instants before launching it in May. Possible fines go up to 10% of global revenue. Not AI, but a reminder that regulators now read Instagram's feature list closely.
- Agents need ID cards. Genesys CEO Tony Bates put the Sierra protocol in one line: brands need a trusted way to know whom an agent represents and what it's allowed to do. Expect that question to come up in every AI tool sale from here on.
The thread through today's news is that customers' AI agents will soon ask your shop the same questions your customers already ask, and they'll want fast, accurate answers around the clock. If those questions are drowning you in DMs right now, that's exactly what Vardast is built for: it learns your products and answers your customers on Instagram, WhatsApp, Telegram and your website, day and night.