Technique

Send a shopping agent through your store before Muse's new model does

Meta's next Muse model is trained to try another route when a website stops it. Before it reaches your product pages, run real errands through an agent yourself and fix what it trips on, in six steps.

22Atlas

Product studio

17 October 20266 min read

A shopping agent can only filter on what it can read. Put the delivery date, the minimum, and the stock status where it can read them.

Yesterday's article was about your agents on other people's websites. This one is the other direction: other people's agents on your store.

On Oct 12, Meta released Muse Wonder, which it calls a frontier model for personal and professional agents. It's available now through Meta's Model API and Muse Code. Meta says Wonder "will soon power experiences in Muse, and will roll out first to subscribers." It gave no date. CNBC reports that Muse app users will need a $20 or $100 monthly plan to get it.

Muse already shops. CNBC walked through how it searches retail sites and shows an approval card before it buys, and TechCrunch reported on agents failing at sites that block them, some on purpose and some by accident. Read those for the basics. What changed this week is the model behind it.

Built not to give up

Meta describes a feature it calls Active Focus as "trying alternative approaches, working through blockers, and adapting to new information." It adds that the model "checks in before taking important actions." Meta AI chief Alexandr Wang told CNBC: "We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time."

Those are Meta's descriptions. Meta's benchmarks are its own, and its full safety report is still to come.

The practical point for a store is simple. If your size picker, bot check, or shipping estimate stops today's agents, the next one is designed to look for another way through. In our 10/16 article, Anthropic described that same persistence as the reason its test agents worked around restrictions on live sites. Meta presents it as a feature. You want your store to be the place where the easy path and the correct path are the same path.

CNBC also reports that some users complained on social media that Muse accessed personal messages without authorization. Meta disputes that. Wang said some users "may have forgotten" authorizing those communications earlier in a conversation.

How Meta says Muse reaches your store

The clearest description came on the same day's episode of Cheeky Pint, the podcast hosted by Stripe co-founder John Collison. Disclosure first: Stripe is a Muse payment partner, Mark Zuckerberg praised the Stripe integration on the episode, and Collison notes on the episode that Stripe has agreed to acquire OpenRouter, which serves Meta's models. Treat the business claims below as Meta's.

Zuckerberg said Meta "trained into the models very good browser use" so that "no one has to do anything to integrate with Muse." Meta is also "going down the list of the most important partners" to offer direct API access, which he said means "you don't have an agent crawling your website." He said many more businesses are asking to integrate than are trying to block it. He gave no counts, so we're not repeating any.

Two of their examples are worth noting. Wang said a shopper can tell Muse "I need something by Wednesday," and the agent "will only look for things that could get to me by Wednesday." Zuckerberg said his own Muse added a watermelon to an Instacart order because the basket wasn't enough to get delivery, and asked him first. He also said "agents are effectively going to have accounts."

Those are anecdotes, not a Meta spec. But they show what an agent reads on your pages: dates, thresholds, and whether it can sign in.

Six steps

Run three real errands through an agent.
Put the facts it filters on in plain text.
Check your bot wall against a customer agent.
Pick your route on purpose.
Decide what a persistent agent may do at a wall.
Watch orders for agent fingerprints.

1. Run three real errands

Use Muse on a paid plan if you have one, or any personal shopping agent you can get. Give it constraints a customer would: arrives by Friday, under a price you pick, a specific size and color. Screen record each run.

Write down every place it slows, loops, or gives up. Common ones are pop ups, size pickers built from images, a "verify you are human" step, and shipping estimates that only show at checkout. Stop each run at the approval card. You're testing the path, not buying from yourself.

2. Put the facts it filters on in plain text

Wang's Wednesday example is a filter on delivery date. Zuckerberg's watermelon is a basket topped up to clear a threshold. Our inference: an agent can only apply those rules to facts it can read.

So show the delivery by date, any free shipping or order minimum, stock status, and the returns window as text on the product page and in the cart. Not only in a banner image, and not only on the last checkout screen.

3. Check your bot wall against a customer agent

TechCrunch reported that Walmart, a Muse partner, said failed purchases weren't intentional. Agents were getting stuck on a button that asks the visitor to prove they're human. Look at your CDN, bot protection, and CAPTCHA settings and find out what a customer's agent sees. Scraper defense and customer agents are two different decisions. Make both on purpose.

4. Pick your route on purpose

There are three. Browser, which Zuckerberg says needs nothing from you. Direct API, which Meta is offering to partners from its own list. Or blocking, which Amazon chose. There's also an open standard in progress: Meta, Sierra, and partners announced Personal Agent Protocol, and Sierra says it plans to publish the v0.1 spec "later this month."

Write down which route you want and who on your team will ask Meta or your commerce platform about it. If you sell on Shopify, also run a Shop MCP test buy on your own store to see what that route looks like from the agent's side.

5. Decide what a persistent agent may do at a wall

Coupon fields, minimums, out of stock items, substitutes, and account sign up are where an agent built to keep going will keep going. Make the right path easy and the wrong path a clear stop. A written substitute policy, a guest checkout that works, and a coupon field that says plainly when a code doesn't apply all count.

This is our advice, not Meta guidance.

6. Watch orders for agent fingerprints

Look for low value add ons that clear a threshold, missing loyalty IDs, and new accounts opened only for an agent to use. Tag them so you can see whether agent orders return, cancel, or get disputed differently from the rest. Keep this short. Its job is to tell you what to fix in step 2 next month.

This week

Run the three errands. Fix the first two things that stopped the agent. Then write one line on which route you want and who owns it.

Tell us which jobs you want off your plate. We install Always On Agents in your own agent account. We still take on build work too. Start your build or write hello@mail.22atlas.com.

Linked sources

Get our Articles and posts by email.

Up to one email a day with new Articles and posts from 22atlas.com and sean-weldon.com, plus occasional news about 22Atlas services.