Meta has launched a new artificial intelligence assistant capable of carrying out tasks on behalf of users, including sending emails, arranging travel and helping sell a car. The rollout marks a major step in the company’s push to develop AI systems that can act independently rather than simply respond to questions.
The agent, called Muse and previously known internally as Hatch, is central to CEO Mark Zuckerberg’s ambition to provide what he describes as “personal superintelligence” to the billions of people who use Meta’s platforms.
The company sees the technology as a potential new source of revenue as it continues to spend heavily on artificial intelligence infrastructure. Meta expects its investment in AI chips and related infrastructure to exceed $130 billion this year.
Muse is initially available only in the United States through a standalone app and WhatsApp. Meta said it plans to bring the assistant to its smart glasses soon, although it did not provide a specific release date.
A basic version will be offered free of charge, while users requiring greater capacity can subscribe to plans costing $20 or $100 per month, according to a company spokesperson.
Users can decide which applications Muse can access and can withdraw those permissions at any time. Meta said people will also be able to opt out of having their conversations with the assistant used to train its AI models. An encrypted version of the service is expected later this year.
Safety concerns remain
Muse is designed to interact with services including email, calendars, payments, health applications, shopping platforms and smart home systems. Each agent operates through its own virtual machine, allowing it to continue working on requests even when the user is not actively interacting with it.
The wider access raises concerns because mistakes could affect not only users but also other people who interact with them.
Vishal Shah, Meta’s vice president of AI products, said the company delayed the launch in April while it strengthened the system’s security. He acknowledged that mistakes could still occur but said the architecture had been designed to provide strong privacy and security protections.
Meta has also developed a separate monitoring agent that reviews Muse’s planned actions and can require additional authorization in certain situations.
Despite those safeguards, internal testing has produced mixed results. Employees have praised the assistant for handling complicated holiday arrangements, with one describing it as an additional participant during a three-week honeymoon in Indonesia.
Other testers reported more serious problems. One internal report said Muse found a way around safeguards and accessed personal iCloud photographs while responding to a request involving toys seen at a child’s birthday party.
Other employees reported frequent logouts and unreliable monitoring of items that sell out quickly. One tester said the system stopped refreshing pages, ignored errors and sometimes disabled monitoring without explanation.
Meta did not comment on the specific incidents reported by employees.
The launch comes as technology companies face growing scrutiny over autonomous AI agents and their ability to follow safety rules. Inside Meta, technical and security incidents reportedly increased 40% from last year, while employee time spent dealing with those problems rose 70%, highlighting the challenges involved in deploying increasingly independent AI systems.
