MAIN CONTENT
An assistant that does the work
When Meta unveiled Muse in early September, it presented the product not as a chatbot but as a personal agent able to act in the real world: scheduling meetings, replying to messages and turning long-term goals into concrete plans. Chief executive Mark Zuckerberg has pitched it as something closer to a helpful friend than a search box.
The pitch worked. Muse climbed rapidly up the app store charts, attracted enormous numbers of sign-ups and pushed competitors to respond, with OpenAI launching a rival tool called Dots. Yet its popularity also gave researchers the chance to look closely at how such an agent operates behind its friendly interface.
“It doesn't just answer questions, it actually does the work.” — Meta
What the code reveals
Experts persuaded the system to export a copy of itself, allowing them to examine the files that make it work. Their findings, published on a website run by the Future Society Hub at the University of St. Gallen, show that Muse runs an hourly loop that creates and updates a page for each person and group tied to its user. These pages can include:

















