Zuckerberg unveils new Muse AI agent capabilities at Meta Connect

Mark Zuckerberg opened Meta's annual Connect conference in Menlo Park on Wednesday and made clear that Muse, the personal AI agent the company launched in early September, sits at the center of its consumer strategy. The agent runs on Muse Spark, a multimodal model built for agentic work, and can plug into email, calendar and other services. Zuckerberg framed the vision in sweeping terms: within a few years Muse will become a "personal superintelligence" used by billions, and the business model will rest on a small commission on transactions the agent executes on the user's behalf, after a period of free use with a generous token allowance.

The most visible announcement is a digital avatar named Jolly, which will gain a face, body and voice through a new model called Muse Realtime Avatar. Users will be able to hold a real-time video conversation with the agent, assign tasks and ask questions — a step that echoes the company's metaverse ambitions, this time in a more personal and immediate form. At the same time, Muse is coming to Meta's smart glasses: wake-word activation, a personalized workout companion, meal logging, appointment booking and purchasing products the user sees. The agent will run in the background and return with results, though Meta cites only "the coming months" as a target window.

On the desktop side, Muse is arriving on Mac with the ability to operate any application on the desktop, similar to Anthropic's Computer Use and OpenAI's Operator. Alexandr Wang, Meta's head of AI, demonstrated on stage that the agent can be left to run a small business or finish chores while the user steps away from the machine. A complementary addition is a dedicated email address for each agent, allowing it to communicate directly with external services and conduct correspondence on the user's behalf.

The announcements are broad in scope, but Meta published no benchmarks for Muse Spark, disclosed no rate limits or precise pricing for future commissions, and committed to no firm launch date for any of the new capabilities, including the glasses and Mac integrations. Given that the product is only weeks old, the gap between the on-stage demo and day-to-day availability remains wide. The test against more established agents will not come from conference promises but from the ability to execute complex tasks in sequence without supervision — and that has yet to be proven.