Mona Sarantakos, head of product for Muse, told Lenny's Podcast how Meta gave its personal agent access to email, calendars and purchases and then held launch to block remote browser attacks.
Why did Meta need seven months to ship when rivals needed weeks?
Sarantakos said trust forced the delay because people will not reuse an agent that fails them once.
Muse is Meta's personal agent that acts from a short prompt across inbox, calendar and web to assemble tasks like a school week brief.
Sarantakos said Friedman pinged a small team around the end of January about OpenClaw and her weekend test on a Raspberry Pi produced a Monday school brief from almost no prompt.
She described an engineering led build with no formal reviews where the team passed a phone to Mark Zuckerberg for feedback while Nat Friedman shaped a Sentinel backstop outside the core agent and Alexandr Wang pushed the surprisingly agentic Muse Spark model.
Reliability, safety and security set the bar because benchmarks felt insufficient, new browser access opened fresh vectors and expanded dogfooding continued until Sarantakos said the team had avoided a catastrophic failure.
Workplace teams that let agents touch shared inboxes must now budget for Sentinel style guardrails and evals, a shift that fits Meta Superintelligence Labs as described in Taskade history as created by a June 2025 memo under Alexandr Wang, and Sarantakos said to watch whether Muse escapes its app trap into glasses and charm devices.
