Meta's new AI agent can email, book, and buy things on its own -- checking out through a card it can't actually see

Meta launched Muse on September 8, a personal AI agent the company calls its first built to act rather than just converse: send an email, book travel, fill out a form, negotiate a price, and complete a purchase, then keep working after the app is closed. It runs on Muse Spark, which Meta calls its most capable model yet for real-world agentic work, and arrives as Meta's answer to a wave of purchasing-capable agents from OpenAI, Google, and Amazon this year.

The name overlaps with, but is distinct from, Meta's other recent "Muse" releases -- Muse Glimmer 30B, an open-weight agent model, and Muse Code, a coding-focused product -- both launched in August. This is the consumer-facing personal agent the Muse Spark model line was built toward, packaged as a standalone app and chat interface rather than a developer tool.

The mechanics are the actual news here. Muse runs inside a dedicated Muse Secure VM isolated from Meta's other systems, and every outbound action -- sending a message, submitting a form, completing a checkout -- has to clear a second, separate agent Meta calls Sentinel before it reaches the internet at all. -- Sentinel has to sign off before any of it goes live. Checkout itself runs through Stripe's Link, which generates a one-time-use card rather than handing Muse a real number -- Meta says Muse "has no visibility into people's passwords or payment methods."

“Nothing Muse does reaches the internet unless the Sentinel approves it.” — Meta Newsroom, “Introducing Muse,” Sept 8, 2026

Meta says the tradeoff for handing Muse real tasks is real limits on what it does with what it learns: conversations and VM data aren't shared with Meta's ad systems, users can opt out of letting interactions train Meta's models, and an upcoming "Muse Confidential VM" is meant to encrypt a user's entire VM with a key even Meta itself doesn't hold. None of that is available at launch -- it's a stated roadmap item, not a current feature.

Meta's own examples are deliberately mundane rather than futuristic: Muse booking a tennis lesson, filling out a school permission slip, or turning a saved Instagram recipe reel into an actual grocery list at checkout. That's the actual bet here -- not a dramatic new capability, but a purchasing agent aimed at the small recurring tasks people already do dozens of times a month, where the tolerance for a mistake is low precisely because the stakes of any single task are small enough that nobody double-checks them closely.

That's the pitch. The skepticism attached to it isn't about whether the engineering works -- it's about who built it. Meta paid a $5 billion FTC penalty in 2019 over user-privacy violations and remains the company at the center of the Cambridge Analytica scandal, and TechCrunch's own coverage of the Muse launch frames its central question around exactly that history: whether users will trust Meta specifically with an agent that touches email, calendars, and a real payment method, regardless of what the VM architecture promises on paper.

Muse arrives into a category every major platform has already tried, with mixed results. Amazon says its Rufus shopping assistant reached more than 300 million users and drove nearly $12 billion in incremental sales during 2025, according to the company's own Q4 earnings disclosure -- a self-reported figure, not an independently audited one -- before Amazon folded it into a broader "Alexa for Shopping" assistant this May. Google has pushed a separate Universal Commerce Protocol aimed at letting agents transact across retailers and payment systems generally, rather than shipping one branded consumer agent of its own. OpenAI's own experience is the more direct warning: it scaled back in-chat Instant Checkout inside ChatGPT in March after adoption stalled -- reportedly around 30 Shopify merchants actually live on it by February, with in-chat checkout converting at roughly a third the rate of sending shoppers to a retailer's own site -- and now points users to partner retailer apps instead of completing purchases directly. Muse is launching with exactly the kind of in-agent checkout OpenAI just backed away from.

Mark Zuckerberg's own framing for Muse reaches well past a shopping assistant: he has said he wants "everyone" to eventually have "an exceptionally capable personal agent" running continuously across their life, a vision PBS NewsHour notes critics have called fantastical. What ships today is much narrower -- a US-only, adults-only agent that emails, books, and buys, gated behind a free tier and two paid ones. Whether that narrower thing actually gets used is, per PBS's own framing, still the open question a launch event can't answer. That gap between the founder's stated ambition and the product that actually shipped is worth holding onto: a continuously running personal agent handling someone's whole life is a much larger claim than an agent that books tennis lessons and fills out permission slips, and only one of those two things exists today.