Meta Launches Muse AI Agent With Dedicated Secure Cloud Computer

Meta has launched Muse, a new personal AI agent designed to perform actions and pursue long-term goals rather than solely answering questions. The agent can execute tasks such as sending emails, booking travel, and negotiating bills, operating autonomously after initial instruction and only returning for user approval on sensitive actions. A key architectural innovation for developers is the provision of a dedicated cloud virtual machine for each user, named Muse Secure VM. This isolated environment hosts the AI agent, its browser, and all user credentials, enhancing security and privacy. Muse is initially available as a consumer service in the United States across iOS, Android, and muse.ai, with both a free tier and paid subscription options. While developers cannot self-host Muse, its underlying model, Muse Spark 1.3, is accessible through the Meta Model API and Muse Code, with an open weights release planned for the future. The Muse agent operates through a messaging interface, where users define tasks or objectives. Muse then plans and executes these actions, including browsing the web, filling out forms, and negotiating on the user's behalf. Meta has provided examples of its capabilities, such as maximizing the sale price of a car, reducing monthly bills, and adjusting fitness training plans. The agent also maintains conversational context, allowing it to remember past interactions and user preferences, like converting an Instagram recipe reel into a grocery list or recalling friends' dietary restrictions. Critical steps, such as sending an email or finalizing a purchase, are designed to pause for explicit user confirmation. A comprehensive audit trail is maintained, detailing all actions taken by the agent and its future plans. The AI model powering Muse is Muse Spark 1.3, which was released last week by Meta Superintelligence Labs. This model is specifically engineered for long-horizon agentic work, featuring capabilities such as zero-shot command-line interface (CLI) tool calling, multi-workflow threading, and self-correction mechanisms when processing diverse and potentially messy data sources. Meta engineers have reported that Muse Spark 1.3 demonstrates improved efficiency compared to its predecessor, Muse Spark 1.2, utilizing approximately 20% fewer tool calls and 25% fewer tokens in internal benchmarks. Meta also claims that the model exhibits near state-of-the-art resistance against prompt injection attacks. Developers can access Muse Spark 1.3 for integration into their applications via Muse Code and the Meta Model API, available at dev.meta.ai. The security architecture, including the Muse Secure VM and a component referred to as Sentinel, is highlighted as a significant aspect of the Muse platform, aiming to provide a robust and secure environment for AI agent operations.