Meta chief executive Mark Zuckerberg ordered his top artificial intelligence executives to ship the company’s AI agent, Muse, despite known safety problems, after a 14-person startup began threatening Meta’s position in the race, according to people familiar with an August meeting at the company.

Zuckerberg met Alexandr Wang, Meta’s chief AI officer, and Nat Friedman, its head of AI product, to discuss the threat from Instinct, a small startup whose AI agent was taking off with users. Meta had been developing Muse, an agent that can operate apps and software on a user’s behalf, but had held back its release over safety worries.

A race against a tiny rival

Wang and Friedman were aware of problems found in testing, including one instance where Muse changed a user’s password without permission, two people with knowledge of the meeting said. Fearful of losing its competitive edge, Zuckerberg declared Muse ready to launch. On September 8, Meta publicly debuted the agent.

The decision reflects the difficult balance AI companies are striking between speed and safety as the technology grows more useful and more dangerous at the same time. Zuckerberg has spent billions developing Meta’s AI, but watched for years as rivals such as OpenAI and Anthropic beat the company to market with chatbots and coding tools. With personal agents becoming the industry’s next frontier, the Silicon Valley giant needed a win, and speed won out.

Three weeks after Meta released Muse, OpenAI announced Dots, its own agent. Meta’s bet appears to have paid off so far. As of last Wednesday, more than 6.6 million people had downloaded the free Muse app, with 1.8 million using it every day, according to research firm Sensor Tower. Meta is now spending millions on a marketing blitz featuring Muse’s fuzzy avatar, a character called Jolly, across billboards and television adverts.

Safety hiccups along the way

The launch has not been without incidents. Muse has handed a user’s address to a seller on Facebook Marketplace without permission, and in another case read a user’s private text messages and then referenced what it had learned. A Meta spokesperson disputed that Zuckerberg pushed the release because of pressure from Instinct, saying the company had delayed shipping Muse for several months to get it right.

The idea for Muse began in July 2025, when Zuckerberg published a 616-word memo about “personal superintelligence”, an assistant-like form of AI. He restructured Meta’s AI division around the concept, hiring Wang and Friedman to lead. The two went on a recruiting spree, hiring researchers from OpenAI and Google with nine-figure pay packages.

An epiphany came in November, when Austrian programmer Peter Steinberger released OpenClaw, an open-source AI agent that could code and use a computer on its own. As it took off among Silicon Valley developers, Meta’s executives decided the next generation of consumer product was a personal agent. Friedman ordered 200 Mac Minis for employees at the Menlo Park headquarters to test agents on, and the company went all in on building its agent.