Meta has introduced Muse, a personal artificial intelligence agent designed to do more than answer questions by taking actions on behalf of users and helping them manage both everyday tasks and longer-term personal goals.
The company said Muse is designed to help users manage tasks, support projects and turn broader objectives into personalised action plans. Unlike conventional AI assistants that primarily respond to prompts, Muse is built to interact with online services and carry out tasks with varying levels of user involvement.
Muse is powered by Muse Spark, which Meta describes as its most capable AI model to date for real-world agentic tasks.
The system can perform activities including sending emails, booking travel, completing online forms and making purchases. It can also continue working on longer-term tasks after a user has closed the app, returning when a decision requires approval or when circumstances affecting the task change.
Meta said Muse can use a dedicated browser to interact with online services. This allows the agent to carry out activities across websites while keeping users involved when an action is sensitive or requires a decision.
The agent is also designed to help users work towards broader personal objectives by coordinating their time and resources. According to Meta, Muse can adjust plans as users’ needs change, potentially allowing it to support projects that require several steps rather than a single interaction.
Another feature of Muse is its ability to learn from conversations and remember information that users consider important.
For example, a user could save a recipe on Instagram and ask Muse to turn it into a grocery list. The agent could also recommend a menu for an event, take dietary requirements into consideration and assist with organising invitations.
This approach is intended to make the AI agent more personalised over time, allowing it to use information from previous interactions when helping users with future tasks.
Meta said privacy and security were central considerations in the development of Muse. The agent operates within Muse Secure VM, a dedicated virtual machine in the cloud designed to keep the agent, user data and credentials separated from other agents.
The company has also introduced a separate security system called Sentinel, which controls Muse’s access to the internet and must approve its online activity.
Meta said Muse does not directly see users’ passwords or payment details. Sensitive actions, including sending emails and making purchases, require user approval before they can be completed.
Users can also determine which applications Muse is allowed to access and control the level of permissions granted to the agent.
For purchases, Muse can use Link, the digital wallet from Stripe, through a one-time-use card intended to keep a user’s actual card details hidden during transactions.
Meta said Muse is the first AI agent covered by Link’s purchase protections. The company also plans to add support for Shop Pay and 1Password.
The system is designed to give users visibility into the actions it takes. Meta said users can review an audit trail of Muse’s activities, disconnect connected services and change permissions whenever they choose.
Users can also instruct Muse to forget specific information and opt out of having their interactions used to train Meta’s AI models.
The company said conversations with Muse and data stored within the virtual machine are not shared with its advertising systems.
Meta is also planning an additional privacy feature known as Muse Confidential VM, which it expects to introduce later this year.
According to the company, the feature will encrypt the entire virtual machine using a key held only by the user. Meta said this approach would prevent even the company itself from accessing the encrypted data and conversations stored within the environment.
The development highlights the broader shift in artificial intelligence from systems that generate responses to AI agents capable of performing tasks in the real world.
Rather than requiring users to complete each step themselves, agentic systems such as Muse are designed to understand an objective, interact with digital services and complete multiple stages of a task.
However, the ability to act independently also increases the importance of security controls and user oversight. Muse’s permission system, approval requirements, audit trail and security infrastructure are intended to give users greater control over what the agent can do.
Meta said Muse is initially rolling out in the United States across iOS, Android and the Muse website. Support for AI glasses is also planned.
The company said the service will be available free of charge to most users, while subscription plans will be offered to people who require additional capabilities.
With Muse, Meta is positioning AI not simply as a tool for answering questions but as a digital assistant capable of taking action, managing information and working towards longer-term objectives on a user’s behalf.