Meta's Muse: how MSL designed a personal agent users can trust
机器之心 · wechat · 2026-11-21
A Machine Heart PRO briefing distills an interview with Mona Sarantakos, VP of product at Meta Superintelligence Labs and lead of Muse, a personal agent. Muse was inspired by the OpenClaw wave: Sarantakos tested OpenClaw on a Raspberry Pi with email and calendar access, and with only a few prompts it synthesized school newsletters and parent portals into her kids' weekly schedule — convincing the team that information-overloaded ordinary users need the same thing. Design decisions included chat-bubble output over the popular long-canvas format, a single main session backed by long-context memory (later supporting multiple sessions), and a customizable plush avatar Jolly instead of a corporate logo to preserve the personal-assistant feel. The full piece also covers how the team judges agent trustworthiness, release gates, and evaluation focus.
More from Companies & People
- Viral joke: Jensen Huang threatened to force employees to buy ETH if they missed deadlines — econoar · 2026-11-21
- Beyond Claude wrappers: real opportunity is data-driven video editing for creators — ayushtweetshere · 2026-11-21
- Palmer Luckey: I've never had an original idea — everything comes from 70-year-old sci-fi — CatAstro_Piyush · 2026-11-21
- Behind 'plane bench': the overnight trip to demo Inkling 2 Mini offline mid-flight — simonguozirui · 2026-11-21
- Ex-Google Gemini/Veo PM, now at Meta, invites Chennai AI builders to meet up — realmadhuguru · 2026-11-21
- Take: $ANT's IPO struggles suggest OpenAI's proposed IPO is dead — SumitGup · 2026-11-21