Meta's Muse agent now has a frontier model and your ad account

Meta is putting the model, the agent and the ad account under one roof. On October 12 it released Muse Wonder and Muse Wonder Full, its first frontier models, and said Muse Wonder will soon power the Muse agent. That agent already connects to small businesses' Meta ad accounts, Shopify stores, Stripe and QuickBooks. It also decides which products to show shoppers, using rules Meta hasn't published.

Our position: treat Meta's agent as a new buyer you have to optimize for, and keep a human approval on every dollar it can touch. Meta's own documentation leaves open exactly how much the agent can do inside an ad account. Third-party research suggests the agent's shopping picks lean toward merchants it considers well-known. Neither of those is a reason to panic, but both are reasons to prepare.

What actually shipped

The new part is the model underneath. Meta's announcement of Muse Wonder says Muse Wonder is available to developers through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers. Meta says Wonder will "soon power experiences in Muse," rolling out to subscribers first. Meta's pitch centers on "Active Memory," which carries context across conversations, and "Active Focus," which keeps the agent working toward a goal and checking in before important actions.

TechCrunch reports API pricing of $4 per million input tokens and $20 per million output tokens, which it calls "roughly equivalent in cost to GPT-6 Sol or Opus 5.5." TechCrunch also says Wonder Full won't be served through the API and that both models will power the Muse agent. Meta's own post gives no benchmark scores, and a fuller safety report is still listed as "available soon."

The agent itself isn't new. Meta launched Muse on September 8 with checkout through Link by Stripe. On September 29 it extended Muse to small businesses with 15 named connectors: Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom. Owners can also connect Facebook Pages, Instagram professional analytics and Meta ad accounts "in a few clicks."

What changed this week is how capable the agent's model is. The access it already had is now in the hands of a much stronger model.

The ad account question Meta hasn't answered

Meta's small-business announcement makes one control promise: "nothing publishes, sends, or spends without your approval." That's a useful promise, and also a narrow one.

The page doesn't say whether ad account access is read-only. It doesn't say whether, once you approve an action, the agent can change budgets, bids, targeting or creative, or how big an approved change can be. PPC Land flags the same gap. Approval-gated is not the same as scoped. If a single approval covers a multi-step plan, you are approving the plan, not each dollar it moves.

There's also a structural point that doesn't depend on the fine print. When the company that sells you ad inventory also supplies the agent that recommends how to spend on it, the advice and the vendor are the same entity. That doesn't make the advice bad. It does mean you need your own measurement to judge it.

How the shopping agent picks products

The more consequential story for merchants is ranking. Here the sourcing needs care, because none of it comes from Meta.

Profound, which sells software that tracks how AI agents treat brands, ran 50 test prompts in one retail category from a single account. It found Muse runs two searches in parallel. A browser typically visited 3 domains and returned 4 products from 2 merchant sites. Meta's product catalog returned roughly 60 products from 30 merchant sites. Final answers usually showed 6 product cards: 4 from the browser and 2 from the catalog. Profound calls the sample small and exploratory.

The "well-known merchants first" claim comes from a separate source. According to PPC Land, Profound's researchers extracted an instruction file from Muse on September 25 and shared it with State of Brand, which published an analysis on October 8. PPC Land then reported on it. As reported, the file calls it "imperative" that the top five picks come from well-known merchants or websites. It caps open-web searches at three stores and puts a named brand's own store first when a shopper asks for that brand. PPC Land also reports that only catalog items carrying an "agentic checkout" flag can be bought inside the chat. Everything else goes through browser checkout.

Treat all of this as third-party and unconfirmed. Meta hasn't commented. The file is weeks old and may have changed, and it's unclear how "well-known" is defined. Still, the pattern lines up with Profound's own test, where the catalog supplies most of the candidates and the open web wins most of the slots. For a small merchant, being in Meta's catalog gets you considered, not chosen.

What this means for SMB operators

Two shifts matter.

First, Meta's agent is a buyer with its own preferences. It reads your catalog, your site and your reputation on the open web, then decides what a shopper sees. That's an optimization surface, much like search was. Merchants that are easy for the agent to parse, purchasable in-chat and recognizable as brands have an advantage. Merchants that are none of those may never show up.

Second, the agent's operator access is growing faster than its documented limits. A Muse instance that can see QuickBooks, Stripe, Klaviyo and your ad account is useful precisely because it can connect revenue to spend. It is also an agent with read access to your whole commercial picture. Its write access to your spend is still loosely described.

The right response is neither to ignore Muse nor to hand it the keys. Engage it on your terms.

What to do about it

1. Audit your Meta catalog for agent-readiness. Check that titles, prices, availability and product attributes are complete and accurate. Profound's test suggests the catalog provides most candidates, and reported filters discard listings priced far outside a category's normal range.
2. Find out whether your products can carry the agentic checkout flag. If the reported rule holds, items without it can't be bought in-chat. Ask your Meta rep or check Commerce Manager, and record what you learn, because Meta hasn't published the requirements.
3. Build signals that you're a "well-known" merchant on the open web. Keep your own site fast and crawlable, make your brand store the obvious result for your brand name, and invest in reviews and third-party mentions. If you resell, expect a brand's direct store to outrank you when shoppers name the brand.
4. Scope the agent's ad account access before you connect it. Use a dedicated ad account or a capped campaign budget for any agent-driven changes. Require approval on every spend action, and reconcile changes against your own reporting weekly.
5. Run your own prompt tests. Once a month, ask Muse for products in your category the way a customer would. Log which merchants appear and whether you do. That's your share of the agent's shelf, and right now nobody else is measuring it for you.