Analysis: the real moat for AI agents is the interruption problem, not the model

A DEV Community analysis of three always-on agents — Meta Muse, OpenAI Dots and Anthropic's Claude Duet (research preview since October 21 on Opus 5.5) — argues the differentiator is not model capability but a three-tier permission system and judgment about what deserves user attention. It cites a Muse incident where the agent disclosed a user's home address on Facebook Marketplace.

• Claude Duet runs on its own cloud computer and sends a morning brief
• Payments and signatures need approval; bank transfers and securities trading are blocked
• Over 60% of users disable an ambient agent within two weeks
• Muse incident: agent shared a home address and accepted a lowball offer unapproved