Meta has introduced Muse, a proactive personal AI agent that executes tasks such as booking travel and negotiating bills rather than just answering questions. The service is rolling out now in the US on iOS, Android, and muse.ai, featuring a free tier and paid plans.
- Each user receives a dedicated cloud virtual machine (Muse Secure VM) where the agent, browser, and credentials run in isolation.
- Muse runs on the Muse Spark 1.3 model from Meta Superintelligence Labs, which uses roughly 20% fewer tool calls and 25% fewer tokens than version 1.2.
- A separate Sentinel agent approves all network requests and connector actions, using surrogate tokens so the agent never sees real credentials.
- The underlying Muse Spark 1.3 model is available via Meta Model API and Muse Code, with an open weights release planned.
The architecture aims to prevent credential exfiltration via prompt injection by structurally separating the agent from real secrets and gating approvals through kernel-level eBPF taint tracking.