Meta's Muse: design decisions behind a proactive background agent

Meta's Muse is a personal agent app, powered by Muse Spark 1.3 according to Meta's announcement, that keeps working on a goal in the background, even while the app is closed. In the designers' account, Muse has its own computer with a file system and terminal, so it can write code and build tools a task needs, plus a full web browser for searching, filling out forms, booking and buying. Once a user sets a goal, it works on a schedule and in response to relevant events, then judges whether a result is worth surfacing and messages the user only when something is meaningfully new or needs input. The designers' example is a parent's back-to-school prep, where Muse watched school emails and caught a sports tryout deadline closing in twelve hours.

The design problem that creates is trust in work the user does not watch, and the designers answer it in two ways. The first is visibility: a snippet under Muse's avatar shows what it is working on, tapping the avatar opens a full activity log and the permissions already approved, memory files can be read and edited directly, and a Goals tab shows everything Muse is tracking and its plan. The second is a small set of controls that do not depend on conversation: structured approval cards with accept and reject buttons, and secure storage for credentials. By default Muse browses the web normally but stops to ask before actions that are hard to undo, such as writing an email or making a purchase, and users can make that rule more or less cautious. The designers say they kept those checks to irreversible steps to avoid banner blindness, where people approve every prompt just to dismiss it.

Proactive messages follow the same reasoning. Because Muse can message the user unprompted, the designers set a high bar for interrupting and let users turn proactivity off, down or up. The post presents these as the decisions and lines the team drew for a product it calls early, and it reports no figures on how often approvals fire or how often background work goes wrong. Muse Spark 1.3, the model underneath, shipped earlier with leaner tool use for agent runs.