Meta Launches Muse, an Autonomous AI Agent Powered by Dedicated Secure VMs

Meta introduced its personal AI agent, Muse, on September 8, 2026. [4] Available through a dedicated app, the Muse.ai website, and WhatsApp, the tool is built to carry out actions like managing schedules and shopping online. [5][9] The agent is powered by Muse Spark, which Meta described as its most capable model to date for real-world agentic work. [7] Prior to the public rollout, Meta employees tested the tool internally under the codename Hatch to autonomously operate third-party applications. [9]

Unlike a conventional chatbot, Muse can operate autonomously after the user closes the app. [7] The agent can open web browsers, fill out forms, and negotiate on a user's behalf. [7] Designed for a mass, non-technical audience, Muse maintains a long-term memory to make unprompted suggestions. [7][8] This allows the agent to recall specific details, such as remembering a guest's dietary restrictions before sending dinner party invitations. [8]

To address privacy concerns, the agent runs on the Muse Secure VM, an isolated Linux computer featuring its own browser, CPU, memory, and storage. [5] A separate system known as Sentinel sits between Muse and the broader internet, using time-limited permissions to distinguish read access from write access. [12] This allows the agent to complete low-risk tasks while pausing to seek explicit approval before executing consequential actions. [12] Every action taken by the system also leaves an audit trail. [12]

For financial transactions, Muse integrates with Stripe's Link checkout system. [7] The integration generates a one-time-use virtual card to hide a user's actual payment details, offering purchase protections covering damaged items and price drops. [7] Support for Shop Pay and 1Password is planned for a future update, which will allow the agent to use existing logins. [7]

The initial rollout of the Muse agent is restricted to users in the United States who are 18 and older. [4][11] Users can access a free tier of the product, but those automating a high volume of tasks must purchase a subscription plan. [9] These paid tiers are priced at $20 and $100 per month. [12] Meta has stated that there is no advertising inside the Muse product, and conversations are kept away from the company's advertising machinery. [12]

A confidential version is slated to arrive later this year that will physically prevent Meta from observing activity inside the workspace. [12] Until then, users have the option to opt out of having their interactions train Meta's models. [12] The product was developed by Meta Superintelligence Labs, an AI unit formed roughly a year ago to catch up with competitors. [9] Support for Meta's smart AI glasses will also be added soon. [9][12]

The launch of Muse pushes Meta further into a direct race against competitors like OpenAI, Anthropic, Google, and Microsoft over who controls the underlying AI agent that acts directly on a person's behalf. By branding the technology as a step toward "personal superintelligence," Meta is attempting to transition consumer expectations from simple chatbot queries to multi-step autonomous workflows. The heavy emphasis on isolated Linux virtual machines, audit trails, and time-limited Sentinel permissions indicates that Meta views security architecture as the primary hurdle for consumer adoption of agentic AI. If successful, the integration of Stripe's Link checkout could establish a new standard for how AI agents safely spend consumer funds without exposing real credit card data. What the sources don't address: How developers of third-party websites might alter their bot-prevention systems to accommodate or block autonomous agents like Muse.