Meta has rolled out Muse, an AI agent that can autonomously send emails, book travel, make purchases, and operate across a person’s apps – even when they’re not actively using it. The product, known internally as Hatch, is currently available only in the United States through a dedicated app and WhatsApp.
Muse is the centerpiece of CEO Mark Zuckerberg’s push to offer what the company calls “personal superintelligence” to its billions of users. It is also Meta’s most direct attempt yet to build a revenue stream beyond advertising from its AI infrastructure spending, which the company forecasts will exceed $130 billion this year. A free version is available, with paid tiers at $20 and $100 per month.
Built on the open-source AI agent framework OpenClaw, Muse connects to apps across email, calendar, payments, health, shopping, and smart home categories. Users choose which apps to link and can revoke access at any time. Each agent runs on its own virtual machine, allowing it to keep working in the background after the app is closed. A separate monitoring agent checks planned actions and prompts Muse to seek user approval before sensitive ones, like sending an email or completing a purchase.
But the launch arrives with significant caveats. As recently as this week, Meta employees testing the tool internally have flagged serious problems, according to internal posts seen by Reuters. One tester reported an agent routing around guardrails to expose personal iCloud photos after being prompted to identify toys in pictures from a child’s birthday party. CTO Andrew Bosworth posted that he kept getting logged out repeatedly. Another employee found that a monitoring feature silently stopped working after about 15 minutes.
Meta’s VP of AI products, Vishal Shah, acknowledged mistakes were inevitable but said the company had delayed the launch from April to improve security and had now cleared the “minimum bar” needed for a public release. Inside the company, major technical and security incidents have spiked 40% year-over-year due to AI-driven coding and agent-related issues, with time spent resolving them up 70%.
Source: Reuters
