September 10, 2026, (Inside AI) — Meta has introduced Muse, a personal AI agent designed to perform real-world tasks on behalf of users. The agent can read email, book travel, fill out forms, negotiate, and even make payments using the user's card. It is now available to US users across iOS, Android, and the web, with support for Meta's AI glasses planned later.
The launch marks a shift from chatbots that merely respond to prompts to agents that act autonomously. Muse continues working even after the app is closed, remembers user preferences, and makes unprompted suggestions. A free tier is available, alongside $20 and $100 monthly subscription plans.
Security Architecture Aims to Rebuild Trust
Meta is betting heavily on security to address public skepticism around consumer AI agents. Each user receives an isolated virtual machine for their agent, preventing different users' agents from interacting with each other. A separate system called Sentinel sits between Muse and the internet, carefully distinguishing read access from write access.
Sentinel allows low-risk tasks using time-limited permissions but pauses to ask for confirmation before anything consequential happens. Meta stores user credentials separately, so Muse never sees passwords. Payments are processed through Stripe's Link, with purchase protections included.
The privacy promise, however, remains a policy rather than a guarantee. Meta states that conversations stay away from its ad machinery, and users can opt out of interactions training its models. But those commitments remain revocable by the company itself.
Agentic Shift Faces Recent Stumbles
The elaborate safety scaffolding has a clear reason behind it. Consumer AI agents have had a rough few months publicly. One Meta researcher's own agent experiment deleted files entirely. Meanwhile, booking platforms like Resy warned against automated agents.
Chief AI Officer Alexandr Wang leads the Muse effort. He framed it around Mark Zuckerberg's personal superintelligence vision. Zuckerberg previously predicted billions using personal agents within five years. Muse is the first product making that claim testable rather than purely rhetorical.
The agent runs on a model called Muse Spark, which Meta describes as its most capable model yet, built specifically for real-world agentic work. The company has not disclosed detailed technical benchmarks or independent evaluations of the model's reliability in high-stakes scenarios.
