Meta Muse Is Better at Collecting You Than Serving You

Meta’s Muse is pitched as the friendly personal agent that books, shops, and clears little chores while you do something else. WIRED’s hands-on review makes the sharper story feel less magical: Muse is better at asking for access than proving it deserves that access.

Quick Take

• Fact: WIRED tested Muse, Meta’s new consumer personal agent, after an Instagram cross-promo pointed users toward an AI helper for deals, reservations, shopping, and inbox work. The app is free, connects to sensitive sources like email and bank data, and wraps the whole pitch in a cute mascot interface.
• Why it matters: Personal agents are permission systems with a personality skin. They do not just answer questions. They ask for standing access, remember preferences, take actions, and turn small choices into durable signals.

  That makes the defaults the story. Training on by default is a big default. Memory without a clearly visible all-off control is a big default. Repeat

• Who cares: Consumer-agent builders designing consent, memory, and connector defaults
• Judgment: **Fairly hyped as a polished mainstream personal agent with strong Meta distribution and a useful web-action loop; overhyped if “built with protections” gets treated as “safe to connect your bank, inbox, and identity doc

What happened

WIRED tested Muse, Meta’s new consumer personal agent, after an Instagram cross-promo pointed users toward an AI helper for deals, reservations, shopping, and inbox work. The app is free, connects to sensitive sources like email and bank data, and wraps the whole pitch in a cute mascot interface.

The upside is real. Muse is not just a chat window with errands stapled on. It can run tasks through a virtual machine, search the open web, click through sites, build carts, and ask for final approval before payment. In WIRED’s bakery test, Muse found a San Francisco shop, assembled a sandwich order, handled a cheese-choice constraint, added a note, and routed payment through Stripe. That is the part builders should not wave away. Background web action is getting smoother.

Meta also has the distribution stack most agent startups would kill for. Instagram can surface the pitch. WhatsApp and Messenger can smooth onboarding. Facebook Marketplace gives Muse a natural job: find local deals, message sellers, and nudge users later about items they liked. WIRED cites Sensor Tower figures putting first-week installs above 900,000, which is not proof of durable use, but is proof that Meta can get a consumer agent in front of a crowd fast.

The problem is the bargain behind the help.

WIRED’s review describes an “ideas” tab that keeps pushing users toward deeper personalization. The suggestions include connecting bank balances for savings tracking, scanning the inbox, photographing documents for form work, snapping meals for calorie estimates, and tracking passport or license expiration dates. That is not ordinary app context. That is mail, money, identity, health-adjacent behavior, and household logistics.

Memory is part of the product. Preferences can live in a long-term memory document that users can edit or wipe, according to the review, but WIRED did not find a simple master switch to turn memory off entirely. Training defaults are also central. WIRED reports that Muse users are automatically opted in so interactions can help train Meta’s AI models, with an opt-out available under Data controls by disabling “Help improve our AI models.” Meta says training data is sanitized. A normal user cannot meaningfully audit what that sanitizing catches, misses, or preserves.

Meta’s defense is that Muse was built with protections and controls from day one. The company also says more “confidential” virtual-machine modes are planned, with cryptographic limits on company access inside the sandbox. That may matter when it ships. For launch trust, it is still a roadmap claim, not a day-one protection users can inspect before connecting a bank account or inbox.

Ads are not framed as a direct “send this chat to the auction” pipeline. Still, WIRED reports that Muse activity, such as bookings or Marketplace behavior, may shape what users see later on Instagram. Privacy advocates quoted in the piece read this as familiar Meta gravity: more context, more personalization, more control over attention. The agent layer makes that gravity easier to miss because it arrives as help.

Why it matters

Personal agents are permission systems with a personality skin. They do not just answer questions. They ask for standing access, remember preferences, take actions, and turn small choices into durable signals.

That makes the defaults the story. Training on by default is a big default. Memory without a clearly visible all-off control is a big default. Repeated prompts to connect more apps, balances, documents, and personal timelines are a big product philosophy. The mascot is cute. The data plane is serious.

For builders, Muse is a useful preview of where mainstream agents are headed: browser automation, approval checkpoints, commerce hooks, and deep personalization. It is also a warning that “more helpful” can become a polite wrapper around “more ingest.” If your agent needs inbox, payment, identity, health, or location context, the consent design has to be stronger than a settings toggle most users will never find.

For companies evaluating agent tools, the lesson is even simpler. Do not judge a product only by task completion. Ask what data it pulls, how long it keeps it, whether training is on by default, whether memory can be fully disabled, what audit logs exist, and whether sponsored or platform-preferred recommendations can appear as neutral help.

Who should care

• Consumer-agent builders designing consent, memory, and connector defaults
• Privacy, security, and legal teams reviewing mail, bank, identity, and document access
• Product leaders watching Meta’s distribution advantage across Instagram, WhatsApp, Messenger, and Marketplace
• Creators and small businesses who may soon meet customers through agents rather than search pages
• Users tempted to connect sensitive accounts because a free helper sounds convenient

Bottom line

Muse looks competent enough to matter. That is exactly why the privacy posture matters too. Meta is showing how personal agents can turn errands into a fresh layer of behavioral context. Help is the trailer. Collection is the runtime.

Bandwagon Check

**Fairly hyped as a polished mainstream personal agent with strong Meta distribution and a useful web-action loop; overhyped if “built with protections” gets treated as “safe to connect your bank, inbox, and identity doc

Sources

• Meta's Muse Is Better at Surveilling Than Helping Me
• Meta’s Muse Is Better at Surveilling Than Helping Me (WIRED)

By Sean Smith · AI Bandwagon