Meta has launched Muse, a personal AI agent designed to carry out work rather than simply answer questions.
Muse can open a browser, complete forms, negotiate, send emails, book travel, and advance longer-term projects after the user closes the app. It returns when circumstances change or approval is needed for a sensitive action such as sending a message or making a purchase.
The agent runs inside a dedicated cloud computer called Muse Secure VM. Credentials are stored separately so Muse can use them without seeing passwords or payment details. A distinct Sentinel agent reviews outbound activity, while users receive granular app permissions and an audit trail of completed and planned actions.
Muse is powered by Meta’s Muse Spark model, but the important development is the complete product architecture around it. Meta is combining persistent memory, background execution, browser access, payments, permissions, and human approval checkpoints in a consumer-facing agent available through both WhatsApp and a standalone app.
The initial rollout is limited to adults in the United States across iOS, Android, and the web. Most usage is free, with subscriptions for heavier use. Meta says interactions can be excluded from model training and are not shared with its advertising systems.
Why it matters
Muse shows what the transition from chatbot to practical agent requires. Model intelligence is only one component; builders also need isolated execution environments, credential protection, policy enforcement, permissions, auditability, and clear intervention points.
Its appearance inside WhatsApp is especially significant. Persistent agents may reach mainstream users through familiar messaging interfaces rather than specialized automation software.