Meta has introduced Muse, a proactive AI agent built to handle long-running personal tasks rather than wait for one prompt at a time. It can keep working while its app is closed, respond to schedules and relevant events, and decide whether new results warrant a notification.
Muse runs on its own computer with a file system, terminal and web browser. It can write code, build tools, navigate sites, fill out forms, complete transactions and create documents, PDFs, web pages and dashboards. In one example, it monitored school emails and a district website, added calendar dates, prepared a supply cart, booked dinner and warned about a sports tryout before registration closed.
The product centers on one persistent conversation. Users can interrupt Muse or submit several tasks before it finishes the last response, while memory retains context and side chats separate complex projects. The agent can initiate messages without a new prompt, with controls to reduce, increase or disable that behavior. Users can also choose its avatar, name and communication style.
Meta has paired that autonomy with visible controls. A snippet under the avatar shows active work, while a full activity log lists actions and approved permissions. Users can inspect and edit Memory files, and the Goals tab tracks tasks and plans. Approval cards provide accept or reject controls, credentials use secure storage, and human review s critical actions such as writing an email or making a purchase. Browsing defaults can be made stricter or looser. Muse can also deliver interactive Artifacts, such as itineraries or live dashboards, and suggest ideas based on goals, patterns and prior conversations.
Mona Sarantakos and Christine Awad, who say they have spent nearly a decade building products together, describe Muse as their most personal and challenging project. The Meta team says the goal is to give people time back while making background work inspectable and keeping hard-to-reverse actions under supervision. Meta is inviting people to try Muse, though its announcement does not specify supported platforms, regions or access limits. The company calls this an early starting point and plans to keep revising the product around usefulness and safety.