Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos

What It Covers

Mona Sarantakos, VP of Product at Meta's Superintelligence Labs, details how a small team built Muse from a weekend experiment with OpenClaw into the fastest-growing AI consumer product in history — covering the seven-month development timeline, key design decisions, safety-first philosophy, and what comes next for personal AI agents.

Key Questions Answered

• •Trust-first launch strategy: Muse took seven months from concept to launch — far longer than competitors like Grok or Dot — because the team set internal safety and reliability bars above industry benchmarks. The reasoning: one catastrophic failure with a personal agent that has access to email, calendar, and purchasing would permanently destroy user trust, making recovery nearly impossible. Launching later with a clean record proved more valuable than speed to market.
• •Agent evals architecture: Muse uses 100+ layered evaluations organized into four categories — safety (prompt injection, harm), capability (browser use, task completion), behavioral (intent understanding, follow-up questioning), and use-case-specific (shopping, productivity, health). Safety evals are non-negotiable gates; others allow minor trade-offs. System prompts and behavioral specs must be handwritten by humans — AI-generated versions consistently produce problems.
• •Agent swarms for product insight: The team deployed agent swarms to analyze multi-turn conversation data at scale, generating both qualitative insights and synthetic training data. This replaced dashboard-only metric reviews with production-level examples — e.g., identifying that 10 users got blocked trying to shop a specific site. PMs are required to review swarm outputs to build empathy for user pain, not just track numbers.
• •Activation through word of mouth: The single most effective lever for converting new Muse users is peer demonstration — watching a relatable person scroll through their actual usage in person. Advertising the concept of "personal superintelligence" failed to resonate until users saw concrete, personal examples. The team built an Ideas tab with pre-packaged starter prompts to reduce first-prompt paralysis, and now develops tiered onboarding variants based on estimated prior AI exposure.
• •Design principle: personal over corporate: Choosing chat bubbles over a large canvas AI interface — and keeping the Jolly avatar despite repeated internal pressure to replace it with a corporate logo — stemmed from a core product principle: the agent must feel like it belongs entirely to the user, not to Meta. Establishing this principle early made downstream decisions (avatar customization, brand minimization) logically consistent rather than arbitrary.
• •Small team, high PM bar: Four PMs built the entire Muse product pre-launch, with a core team under 100 people total. The team deliberately avoided adding PMs to zero-to-one exploration phases, bringing them in only when features approached launch readiness. The current top criterion when hiring PMs is product taste — the ability to represent consumer needs against technically capable engineers who can now prototype and ship faster than ever before.