Meta Moves In Downstairs

The consumer agent era stopped being a pilot program on Tuesday. Meta moved Muse — the assistant it has been testing with a slice of its apps — into general availability, and the shape of the thing is the announcement: an agent that does not run on your phone or your laptop but in a dedicated “secure virtual machine” in Meta’s cloud, with its own browser, holding your sessions while you are elsewhere. Free to start. $20 a month at Power, $100 at Maximum. Three billion people one menu away from handing a company their errands, their logins, and their judgment calls.

The architecture is the tell. A phone assistant answers when asked; Muse’s secure VM persists — it books, compares, drafts and revisits across hours while the user does something else, which is the entire point of an agent and the entire problem. Meta’s framing is that the VM is a privacy feature: credentials and sessions live in a sandboxed cloud computer rather than spread across a personal device. The framing writes its own counterargument in the same sentence — the agent that holds your keys is not the agent you can watch. TechCrunch’s launch-day headline put the question where the company could not: will consumers trust it?

The pricing answers a different question. Free-to-start with a card on file, Power at $20, Maximum at $100 — the ladder that OpenAI and Anthropic built for knowledge workers, aimed instead at the default apps of three billion people. Meta is not selling a better model; Muse runs on house silicon and house weights. It is selling presence: the agent that is already installed, already logged in, already in the group chat. Distribution has decided platform shifts before — and Meta owns the floor the shift would happen on.

The Trust Ledger

The trust question is not rhetorical for Meta specifically. The company’s last decade of consumer AI-adjacent history — feed algorithms tuned for engagement, the Cambridge Analytica inheritance, a metaverse pivot that consumed billions in goodwill — is precisely the record a user reviews before handing an agent their calendar. Meta’s counter is structural: the secure VM, published safety documentation, and the arguable claim that a sandboxed agent is more auditable than a dozen browser extensions with the same permissions. The arguable claim will get its argument in public soon enough.

The competitive map underneath is the quiet story. Muse general-availability lands on the same day the frontier labs are selling agents to enterprises at $200 a seat and OpenAI is measuring its research automation in agent-workdays. Meta is proposing the mass-market branch of the same technology tree: less capable on a benchmark, more capable in a household. If the consumer agent market forms — really forms, with retention and habit, not novelty — it will not form first among developers with API keys. It will form among people who already open Instagram before their eyes focus.

The frontier labs built agents for the people who build agents. Meta built one for everyone else — and the everyone else is three billion strong.

What to watch: retention curves, not launch-day traffic — agents live or die on whether anyone comes back in week three. And watch the first trust incident, because there will be one: the first purchase made wrong, the first message sent without permission, the first credential held in a secure VM that ends up somewhere it should not. The enterprise agent market absorbed its incidents in post-mortems. The consumer market will absorb its first one in a news cycle, and the size of that news cycle will tell us how the next decade of consumer AI is governed.