The agent is the new storefront

2026-09-10

Meta launched Muse this week, a personal agent that lives on its own virtual computer, browses the web, and completes tasks for the user: it books the restaurant, fills the form, buys the thing. It reached the top three of the US App Store on its second day. Early usage runs at roughly ten times what the test groups showed.

The product is free for most people. There are paid tiers, but Meta has said plainly that the money is meant to come from a take rate on transactions that pass through the agent, not from subscriptions. That sentence matters more than the launch itself.

Distribution, not models

The last two years of AI commentary were about who has the best model. Meta's own models now sit close to the frontier, but that is not why this launch counts. It counts because the model is wired into WhatsApp, Instagram, Messenger and Facebook, which together reach about four billion people and more than two hundred million businesses. A good model with no distribution is a research paper. A near-frontier model with four billion users is a channel.

On the business side, more than a million businesses were already using Meta's business agent on WhatsApp and Messenger every week by mid-year, answering questions, recommending products and booking appointments. Enterprise access is billed by the token at a few cents per message. Meta has said the next step is an outcome-based auction, which is the ads business applied to agents.

What changes for a brand

For the retail and cross-border work we do in Southeast Asia and Brazil, the practical implications are concrete.

The customer may never see your storefront. A shopper asks the agent for a product, the agent compares, picks and pays. Your listing, price and delivery promise are read by software, not by a person.
The platform owns the transaction layer and will charge for it. A take rate on agent-led purchases is a new line item between the brand and the buyer, on top of the ad spend already going to the same company.
Catalogue quality becomes the marketing. Structured product data, accurate stock, honest delivery times: these are what an agent ranks on. A pretty page it cannot parse is invisible.
Agent-to-agent is the end state. Meta expects merchants and enterprises to run their own agents that negotiate with consumer agents. The brands that get there first set the terms.

The cost of being the channel

None of this is cheap for the company building it. Meta has said it will maximise compute capacity through 2026 and 2027, with reports of seven gigawatts this year doubling next year. Street estimates have its capital spending above two hundred billion dollars in 2027, pushing free cash flow negative for two years. That is the price of owning the storefront of the agent era, and the whole industry is paying some version of it.

Our read

We are sector-agnostic, but we are not channel-agnostic. The brands we take into new markets will increasingly be sold by agents they do not control, inside apps they do not own, to shoppers they never meet. The work is to make the product legible to the agent, to price for the take rate before it arrives, and to keep a direct relationship somewhere the platform cannot reach. The launch this week was for consumers. The bill is for the brands.