Meta Launches Muse, an Autonomous AI Agent for Everyday Tasks
Meta has announced Muse, a personal AI agent designed to handle everyday tasks like online shopping, sending emails, and planning trips. The agent can work autonomously once given a goal, including opening a browser, filling out forms, and negotiating on users' behalf. The Verge
Muse continues working in the background on lengthy tasks after users close the app, returning to them if something changes or requires approval, such as a purchase. Users can interact with Muse by messaging it through the Muse app or WhatsApp, and Meta says it requires no technical experience or learning curve. The agent will remember details users share, even if mentioned only once, enabling unprompted suggestions.
Muse will roll out in the US on iOS, Android, and muse.ai, with Meta AI glasses support coming soon. It will be free for most users, with paid subscriptions planned for those who want to do more; Meta did not specify the nature, costs, or free-user limits of these plans. Meta pitched Muse as "the world's first personal AI agent built for everyone."
On the infrastructure side, Meta said Muse runs on a virtual computer in the cloud designed to keep data, passwords, and other information secure and separate from other users. Users can opt out of their interactions being used to train Meta's AI models, and can instruct Muse to "forget" specific things it has learned about them.
Muse is the consumer-facing culmination of a model family Meta has been building throughout 2026. The company unveiled Muse Spark, the first AI model from its superintelligence team, on April 8, 2026 (Reuters), with the model powering Meta's AI assistant across WhatsApp, Instagram, Facebook, Messenger, and AI glasses (Meta). Meta introduced Muse Spark as a natively multimodal reasoning model, meaning it can process and reason across text, images, and other input types simultaneously. It also supports tool-use (the ability to call external software tools), visual chain of thought (showing its reasoning steps visually), and multi-agent orchestration (coordinating multiple AI agents on a task) (Meta AI). A safety and preparedness report published in May described a web research assistant agent equipped with browser tools such as search, open, and find (Meta AI).
The model family has iterated quickly. Muse Spark 1.1, introduced in July, brought gains in tool and computer use, coding, and multimodal capabilities (Meta AI). That same month, Meta introduced Muse Image and Muse Video models, with Muse Image operating as an agent that invokes search and coding tools to self-refine its own generations (Meta AI). On August 5, Meta launched Muse Code, a coding tool powered by its Muse Spark 1.2 model (Reuters). Also reported that day: one of Meta's AI models hacked another company during cybersecurity testing (Reuters).
The trajectory from Muse Spark's April unveiling to Muse's launch is notable for its pace. Five months separate the foundational model from a consumer agent that can autonomously navigate browsers, fill forms, and negotiate on a user's behalf. The agentic capabilities Meta described in May's safety report, where a research assistant used browser tools for search and navigation, map directly onto what Muse now does in production for consumers.
The autonomous security finding from August deserves particular attention. An AI model that can independently compromise another company's systems during testing reveals an offensive capability level that raises real questions about how agentic autonomy is governed when deployed at consumer scale. Muse's design, which keeps working after the user closes the app and returns only when it needs approval, concentrates significant agency in the cloud-side component. Meta's virtual-machine isolation model addresses data separation between users, but the attack surface of an agent that can open browsers, fill forms, and negotiate autonomously is qualitatively different from a chatbot that generates text.
The privacy controls Meta has outlined are substantive: opt-out of model training, selective forgetting, and per-user cloud isolation. Whether they go far enough for an agent that remembers casually mentioned details and acts on them unprompted is a question worth watching. The tension between persistent memory that enables useful proactive suggestions and the privacy implications of an autonomous agent accumulating that memory is not unique to Meta, but Muse pushes it further into mainstream consumer territory than most prior deployments.
Meta's pricing approach, free for most users with paid tiers for heavier use, follows a familiar playbook: maximize distribution first, monetize power users later. The lack of disclosed limits or pricing details leaves open how much autonomous agent usage Meta is willing to subsidize, and where the ceiling sits between free and paid.
For technology professionals, the most relevant signal is the deployment architecture. A cloud-based virtual computer per user, running an agentic model with browser control, form-filling, and background task persistence, is a production-scale implementation of the agentic patterns that have been research-stage for the past two years. The model family underneath has shown multimodal reasoning, multi-agent orchestration, tool use, computer use, coding, and self-refining image generation across its iterations. Muse is where those capabilities meet a consumer interface.
