Meta's Muse AI agent accelerates the race for autonomous software

Meta launched Muse in the United States on September 8, with an official announcement of the personal agent. Muse is an AI agent capable of autonomously performing operations on applications and software. The launch marks a turning point in the race for AI agents. According to a reconstruction by the New York Times published on October 9, the launch was decided at the top of the company despite some critical issues that emerged in internal testing. For founders and investors, it is the signal that the market has shifted. The conversation is no longer about chatbots. These are software programs that act directly in users’ digital environments. And in this market, giants with more than 3.5 billion users are clashing with startups of just over a dozen people.

Muse’s accelerated launch and the Instinct threat

In August, Mark Zuckerberg reportedly summoned Alexandr Wang and Nat Friedman, according to the New York Times. Wang is in charge of artificial intelligence at Meta. Friedman heads AI products. The topic was an unexpected danger: Instinct. The 14-person startup was gaining attention with a competing agent. During the meeting, Zuckerberg reportedly argued that Muse was ready for the market. Testing, however, had flagged unexpected behaviors, including changing a user’s password without authorization.

Meta prioritized speed to market. A company spokesperson disputed the account that the launch had been accelerated by competitive pressure. The spokesperson recalled that the product had been delayed for several months, to make sure the company got it right.

Muse’s first commercial results

The early data seem to vindicate Zuckerberg’s strategy. According to Sensor Tower estimates cited by the New York Times, Muse surpassed 6.6 million downloads, with 1.8 million daily active users. Meta has also given the agent a soft-looking mascot, Jolly. The real advantage, however, remains distribution, with the more than 3.5 billion people who use the company’s platforms every day as a potential pool.

OpenAI has entered the field too: on September 29, three weeks after Muse’s debut, it announced Dots, a competing agent.

From personal superintelligence to technological independence

Muse stems from the “personal superintelligence” vision presented by Zuckerberg in July 2025. It is a digital assistant capable of understanding individual needs. It can also act autonomously in daily activities. To support it, Meta reorganized its AI division. It entrusted model development to Wang and products to Friedman. It launched a recruiting campaign among OpenAI and Google researchers. In some cases, it offered nine-figure compensation packages.

A decisive impetus came from OpenClaw. It is the open source agent of Austrian programmer Peter Steinberger. It is capable of controlling computers, writing code, and using applications autonomously. Friedman even ordered 200 Mac Minis to let employees experiment with OpenClaw in the offices. On the technology side, Muse initially used models from Anthropic: the introduction of Muse Spark, a proprietary model developed under Wang’s supervision, reduced that dependency. Controlling the entire AI infrastructure makes it possible to intervene on performance and safety without external vendors.

Privacy and the risks of autonomous agents

Operational autonomy introduces problems different from chatbots. During testing, Muse reportedly occasionally ignored users’ instructions or directed people to fraudulent websites. After the launch, privacy complaints emerged: one user said Muse had given their address to a Facebook Marketplace buyer without asking, and an Inc. columnist claimed the agent had read their private messages without permission, a claim Meta denied.

OpenClaw has also shown problematic behaviors. Summer Yue is a Meta researcher specializing in security. She recounted that an agent had started deleting her emails uncontrollably. She was forced to physically intervene on the computer.

These episodes make the issue of permissions, human oversight, and agents’ operational limits central when they access personal data or modify external services.

The next battle of artificial intelligence

The Instinct case shows how asymmetric the competition is. The startup was founded by Noah Shinn. As TechCrunch reported on September 28, it raised $1 billion in a Series C round at a $10 billion valuation. It demonstrates that small teams can compete on product with much larger organizations. The difference remains on the resources front: Meta’s computing capacity, capital, and distribution channels are hard to replicate.

The race for personal agents shifts the competition from content generation to the autonomous execution of digital tasks, and development speed can favor new players.

Significant opportunities are opening up for startups, but pressures are growing: safety, reliability, and distribution at scale require substantial investments. The outcome will depend on the ability to turn agents’ autonomy into an everyday service that does not compromise users’ control over their own data.

Sources: