Muse AI agent launched by Mark Zuckerberg's Meta looks at old saved Instagram reels on a recipe and send a grocery list; check what else it can do

Scroll through your saved Instagram reels long enough and you'll probably find a recipe you meant to try weeks ago. Meta says its new AI agent, Muse, can do something about that on its own: spot the saved recipe, work out what's needed, and hand over a ready-made grocery list without being asked.

That's just one of the everyday tasks Meta is pitching Muse for as it begins rolling out in the US this week. But the launch has come with a mixed backstory, with internal Meta staff flagging real security concerns about the tool even as the company pushed it out to the public, Reuters reported.

What Muse can actually do


Muse is built to remember details a person has mentioned only once and act on them without a fresh reminder, Meta said. Alongside turning a saved recipe reel into a grocery list, the company said Muse can suggest a dinner party menu while keeping track of which guests have dietary restrictions, and can carry out day-to-day chores like sending emails, booking travel, and filling out forms on a person's behalf.

Once someone shares a bigger goal, Meta said, Muse builds a plan, coordinates the person's time and resources, and keeps working on it in the background even after the app is closed. It comes back only when something changes or when it needs approval before a sensitive step, such as sending an email or making a purchase.

How it's built to work

Each Muse agent runs on its own dedicated cloud computer, called Muse Secure VM, according to Meta. A separate agent named Sentinel sits on the same machine and screens Muse's planned actions, so nothing it does reaches the internet unless Sentinel approves it or the person gives the go-ahead. Meta said Muse itself never sees a person's passwords or payment details, since any credentials shared with it are stored separately and used without being visible to the agent.

For payments, Muse checks out using Link, a tool built by Stripe, and Meta said it is the first AI agent to carry Link's purchase protections, including cover for damaged or lost items, price drops, and fee-free returns. Each transaction uses a one-time virtual card, so a person's real card details stay hidden. Shop Pay and 1Password support are expected to follow.

People choose exactly which apps Muse can connect to and how much access each one gets, Meta said, and can change or revoke that access at any time. Users can also tell Muse to forget specific things it has learned, and can opt out of having their interactions used to train Meta's other AI models. A more locked-down version, called Muse Confidential VM, is expected later this year, encrypting a person's entire virtual machine, including their data and conversations, with a key Meta says only they will hold.

The story behind the launch, as per Reuters

Internally, Muse was developed under the codename Hatch and is modeled on OpenClaw, an existing open-source AI agent, Reuters reported. The product is priced with a free basic tier for most everyday use, alongside subscriptions at $20 a month and $100 a month for heavier use, Reuters said, citing a Meta spokesperson.

Meta had originally planned to release Muse earlier this year, around April, but delayed it to make the product more secure, Vishal Shah, the company's vice president of AI products, told Reuters. Shah said the extra work helped the agent "cross the threshold" needed for public release, while acknowledging that mistakes are still possible.

Internal testing turned up problems, Reuters found

Despite the delay, Meta employees testing Muse ahead of launch reported mixed results, according to internal posts seen by Reuters. One tester said Muse became "the third participant" on a three-week honeymoon, handling travel logistics smoothly throughout the trip. Others flagged more serious issues, including an instance where an agent asked to identify toys in a child's birthday photos ended up exposing that person's entire personal iCloud photo library, Reuters reported. Meta did not respond to Reuters' request for comment on that specific incident.

Meta's own chief technology officer, Andrew Bosworth, posted internally that he kept getting logged out of Muse and had to sign back in several times within a few minutes, according to Reuters. In a separate case reported by Reuters, an employee who had set Muse to monitor for fast-selling tickets said the tool stopped refreshing the page after about 15 minutes, silently ignored other errors, and at times switched off its own monitoring without explanation.

Reuters also noted that the launch comes as AI agents from several major companies, including OpenAI and Anthropic, have shown similar patterns of unpredictable behaviour. Inside Meta, major technical and security incidents have risen 40 percent compared with last year, driven partly by a surge in AI-related coding and agent issues, while time spent by staff "firefighting" such incidents has climbed 70 percent over the same period, Reuters reported.

Muse is rolling out first in the US through a dedicated app and through WhatsApp, with support for Meta's AI glasses expected soon.