Meta bets on AI agent Muse to catch up in AI race

The Verge frames Muse as Meta’s latest push to put AI in ordinary hands after a multi-billion-dollar strategy overhaul aimed at catching rivals. Once given a goal, Meta says the agent can work alone—opening a browser, filling forms, negotiating—and keep going in the background after the app closes, returning when something changes or a purchase needs approval.

Rollout is U.S.-first on iOS, Android, and muse.ai, with AI glasses “coming soon.” Meta pitches free use for most people and paid subscriptions “for people who want to do more,” without detailing free limits or paid prices in The Verge’s account. Messaging can happen in the Muse app or WhatsApp; Meta claims no technical experience is required and that details mentioned once are remembered for better, even unprompted, suggestions.

Trust is the gating factor. Agents need personal information to be useful; Meta says users can opt out of training use of interactions (default unclear) and can tell Muse to forget specifics. Muse runs on a cloud virtual computer; Meta says it has no visibility into passwords or payment methods, with another agent, Sentinel, policing what reaches the internet. A more encrypted “confidential” VM is planned later this year, plus 1Password and Shop Pay, on top of Stripe Link checkout already available.

Signal Desk’s read: Muse is less a product novelty than a distribution play—Meta’s household reach versus labs whose best agents still skew toward businesses and technical users—and the company knows privacy theater is the price of admission.

The competitive map The Verge sketches is accurate: ChatGPT Work, Claude Cowork, Copilot Tasks, Grok Bot, open-source Moltbot, and Google’s Gemini Spark already crowd the agent shelf. Meta’s differentiation claim is blunt—“anyone can use it out of the box,” “no learning curve,” WhatsApp as a familiar surface. That is a strategy for casual users, not for beating frontier labs on model bragging rights. Powered by in-house Muse Spark, the agent is being cast as the centerpiece of Meta’s catch-up story after years of setbacks.

What Meta under-states is how much baggage rides with that household name. The Verge notes a grueling social-media harms trial, Cambridge Analytica, Discover exposing other users’ prompts and images, a support chatbot that helped hackers take over more than 20,000 Instagram accounts, a short-lived AI clickbait news product that appeared to break Meta’s own AI-content policies, and the “pervert glasses” reputation. Ease of use without trust is just a faster path to credential exposure.

The likelier read is that Meta is racing to own the default consumer agent channel before Google’s Gemini Spark and chat-native rivals lock habits. Sentinel, training opt-outs, forget commands, and a future confidential VM are necessary product hygiene, not proof of rehabilitated culture. Watch whether Meta publishes clear free-tier limits and paid pricing soon; opacity on meters is itself a trust tell when the ask is email, browsing, and purchases.

Meta has spent heavily to reenter the AI race, including a reported $14 billion catch-up push and the Muse Spark model relaunch. Muse’s consumer pitch sits beside enterprise-leaning agents from peers and Google’s similar mass-audience Gemini Spark assistant.