Meta unveils Muse, its first personal AI agent that acts on the user's behalf

Meta has announced Muse, a personal artificial intelligence agent that, according to the company, does not merely answer questions but carries out real tasks on the user's behalf. The core proposition is a shift from a conversational assistant to an agent that manages to-dos, coordinates projects and turns long-term goals into action plans it advances autonomously.

It runs on two technical pieces. On one side, Muse Spark, described as the most capable model Meta has developed to date and geared specifically towards real-world agentic work. On the other, Muse Secure VM, a dedicated cloud virtual machine that hosts both the agent and the data and credentials of the services the user connects. The interaction is designed to require no learning curve: you talk to Muse as if messaging another person, either from its own app or directly in WhatsApp.

As for capabilities, Muse can handle both one-off tasks (sending an email, booking a trip) and broader goals: once the user shares an objective, the agent helps design a personalised plan, coordinates time and resources, and keeps working even after the person closes the app, getting back in touch when there is news or when it needs authorisation. The examples the company offers include selling a car for a higher price, negotiating a discount on a bill or adjusting a training plan as the person's circumstances change. To do so, Muse can open a browser, fill in forms and negotiate on the user's behalf.

A specific section is devoted to payments. Muse can complete purchases through Link, the solution built by Stripe, and is presented as the first AI agent covered by Link's purchase protections: free coverage for damaged or lost items, price-drop protection, free returns and a money-back guarantee on eligible purchases. Link's agent wallet generates a single-use card that hides the details of the user's real card. Meta also announces that Shop Pay will soon be added as an additional payment method, along with 1Password compatibility so that Muse can reuse credentials the person has already saved.

Another pillar of the product is personalised memory: Muse remembers relevant information in order to make suggestions without being explicitly asked. The article gives the example of turning a recipe reel saved on Instagram into a shopping list, proposing a menu for a dinner party or recalling the dietary restrictions of the user's friends before sending out invitations.

Security and privacy take up a central part of the announcement, predictably so given that this is an agent able to act autonomously on real accounts and payments. Muse runs on a dedicated, isolated computer in the cloud that, according to Meta, no other agent can access. On that same machine, but separated at system level, a second agent called Sentinel operates: nothing Muse does reaches the internet without its approval, and Sentinel asks the user for permission when necessary. Meta also states that Muse has no visibility into passwords or payment methods —credentials are kept in secure storage the agent can use without seeing them, including passwords the person types directly into the browser— and that it checks with the user before sensitive actions such as sending an email or making a purchase. The system also offers a full audit log of what Muse has done and what it plans to do.

Access control is granular: the user decides which applications Muse connects to and with what permissions (for instance, whether it can only read email or also send it on their behalf), and can modify those permissions or disconnect a service at any time. There is also the option to exclude one's own interactions from the training of Meta's models, and the company assures that conversations and data hosted on the VM are not shared with its advertising systems. The user can also ask Muse to "forget" specific information it has learned.

Looking ahead, Meta announces that later this year it will introduce Muse Confidential VM, a variant in which the entire virtual machine —including the user's data and their conversations with Muse— would be encrypted with a key only that person holds, so that not even Meta itself could access its content.

On availability, Muse launches initially in the United States for iOS, Android and via muse.ai, with a later rollout planned for the company's AI glasses. Access is free for most uses, with paid subscription plans for those who need more capacity.

The announcement fits into a broader Meta narrative about what it calls "personal superintelligence", presented as one of the most transformative technologies the company expects to develop, of which Muse would be a first step. The article itself links to other previous company announcements related to the same product family —Muse Spark as the base model and Muse Image for image generation— suggesting that Muse is part of a personal agent strategy built on several components developed in stages over the course of this year. It is worth noting that all the information gathered here comes from Meta's own release; there are, in the available material, no independent evaluations of the agent's actual performance, its error rate or its behaviour under mass usage conditions, aspects that will foreseeably only be verifiable once the product is in users' hands.