(Meta's Muse)
What's really inside Meta's Muse — the loop, the memory, the subagents and the guard at the door — in about half an hour.
A free, visual course on how Meta's Muse AI agent works: the plan-act-observe loop, the Muse Spark reasoning model underneath, connectors and browser use, long-term memory, background subagents, and Muse's security design — a Secure VM, the Sentinel guard and stand-in credentials — that stops prompt injection.
Muse books travel, answers email, sells your old stuff and haggles over bills on your behalf — and every personal AI agent is built from the same parts. Knowing how Muse works tells you what to trust it with, and what to keep an eye on.
Muse acts through tools — typed connectors when an app has an API, and eyes-and-mouse on a browser when it doesn't.
The model forgets everything between calls. Muse's memory is a file it writes to — and reads back into its context later.
Close the app and the work continues — on a computer in the cloud, split across subagents, until it needs you.