Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos
· Summarized
Key takeaways
- Muse began when Mona spent a weekend using OpenClaw on a Raspberry Pi and saw a tiny prompt produce a useful personal digest from her email, calendar, and school information.
- The team delayed launch for seven months because reliability, safety, security, and user trust mattered more than matching competitors’ speed; its internal rule was “trust is a must.”
- A consumer agent has to feel personal and user-owned, which is why the team preserved the Jolly avatar instead of replacing it with a corporate Meta logo.
- Shopping, not home automation, became Muse’s unexpected major use case, and the product is intended to help users narrow choices without eliminating the pleasure of browsing.
- Agent swarms let the team analyze long, multi-turn production conversations at scale, turning raw usage into specific product insights and greater customer empathy.
- The strongest activation mechanisms are relatable word-of-mouth stories and giving new users concrete starter ideas so they can get past the intimidating first prompt.
- AI has shifted the PM role toward product taste, empathy, rapid prototyping, and maintaining a high quality bar rather than primarily writing specifications.
- Mona’s long-term vision is for Muse to become an ambient agent in glasses and other devices, while requiring less user babysitting and contributing to users’ health, time, and financial well-being.
Summary
Mona’s background and the unexpected start of Muse
Mona Sarantakos, Meta’s VP of product in Superintelligence Labs, had spent about 10 years at Meta and roughly two years working in AI, primarily on Meta AI. Her team was “zero basing” Meta AI—starting over rather than continuing to optimize a product that was not working as desired—when Nat Friedman messaged a few people around the end of January or beginning of February. He said OpenClaw had changed his view of what personal superintelligence and consumer agents could become and urged them to try it. Mona did not receive a formal project pitch; a short message became an informal invitation to experiment.
A Raspberry Pi weekend revealed the consumer opportunity
Mona could not find a Mac Mini, so her husband helped her set up OpenClaw on a Raspberry Pi. While her children were away at ski racing for the weekend, she connected it to her email, calendar, and home automation and spent the weekend talking with it. On Monday, it assembled a digest from school newsletters and portals—something other assistants could technically do, but which felt magical because Mona had given it only a tiny prompt. She suddenly felt productive and caught up in her personal life, recognized the potential for busy parents, and told Nat, “I’m in.” The team then realized a consumer version had to be safer, easier to set up, and scalable.
Why Muse took seven months to launch
Although a functioning end-to-end product existed by spring, Mona’s team spent seven months reaching a launch bar they considered acceptable. The early phase focused on defining the product and solving basic infrastructure problems such as running virtual machines; later work concentrated on reliability, safety, and security. Browser access introduced new security risks, and each important feature created additional attack vectors to close. The team created its own evaluation standards because existing industry benchmarks were insufficient for an agent with personal context and the ability to buy things. Their internal rule was “trust is a must”: launching with serious failures could permanently destroy users’ willingness to trust an agent.
Dogfooding replaced a conventional prelaunch process
Muse was not simply held back while the team waited for perfection. Meta employees dogfooded it, an external alpha panel used it, and the team gradually expanded the number of people providing feedback throughout the summer. Mona says the product still cannot eliminate every unknown because the industry is so early, but the expanding test groups made the team more confident by launch. The internal vision was captured in an early document called “the claw is the law,” a phrase Nat Friedman used after the team saw OpenClaw’s potential and needed a way to explain to colleagues why this unfamiliar, seemingly risky product mattered.
How Nat Friedman, Alex Wang, and Zuckerberg shaped Muse
Nat effectively acted as part of the small team, pushing the original OpenClaw insight and getting deeply involved in architecture, security, and the Sentinel—a backstop outside the core agent. Alex Wang brought the unexpectedly agentic Muse Spark model and adjusted research priorities around what the product needed; he ultimately pushed the team to launch by saying it was ready. Mark Zuckerberg repeatedly tested prototypes on the team’s phones and reframed Muse around helping people achieve their goals rather than merely completing tasks. That led to proactive “dreaming,” “study,” and the Goals tab, as well as a focus on making personal superintelligence broadly accessible and scalable.
Jolly survived because a personal agent needed a personality
Lead designer Alex Cornell initially created Jolly Bot as a temporary creature in Muse’s prototype, borrowing the name from his nickname for his daughter. The team kept adding motion, headphones, a laptop, and eventually a system for users to create their own avatars, even as people repeatedly insisted that a Meta product could not launch with a stuffed animal at the top. Jolly nearly disappeared in favor of a corporate-style logo, but Zuckerberg and the team resisted because a personal agent should not feel like a corporate entity. Mona says the avatar made the product feel like it belonged to the user, not Meta, and later became iconic partly through Alex Wang’s public enthusiasm.
Shopping became Muse’s surprise leading use case
The team initially prioritized personal productivity, health, and personal finance, with home automation also expected to be important. User research and dogfooding instead revealed that shopping was a major use case, forcing the team to build new shopping evaluations and prioritize it. Engineering leader Christine Wad became the team’s most enthusiastic shopping dogfooder, to the point that Mona jokes she needed to take away Christine’s credit card. Mona also began using Muse to shop across multiple stores; instead of bulk-ordering five pairs of sweatpants from Nike, Muse selected items from several retailers, producing better variety but also seven separate packages at her house.
Muse is meant to narrow choices, not eliminate browsing
Asked how agents will change retail and advertising, Mona argues that shopping has two different modes. Sometimes users want a utilitarian answer—such as finding toilet paper quickly—but people also enjoy browsing, discovering outfits, and being inspired. Muse should help define the option space and make difficult trade-offs when users are paralyzed by choices, while preserving the browsing experience rather than replacing it. Mona expects consumer agents and business agents to communicate, allowing merchants to provide information directly to an agent and potentially making shopping more useful on both sides. She says the industry is still figuring out this model because no master plan existed before agents arrived.
Agent swarms turned production conversations into product insight
Muse’s team created “agent swarms” that examine production conversations, metrics, analytics, and user feedback, then return structured summaries of what users are doing and where they are struggling. The swarms can also generate synthetic data. They were especially valuable because Meta’s earlier assistant data was dominated by short, often single-turn interactions, while Muse required datasets for long-running, multi-turn tasks such as meal planning. When Mona suspected a problem might be unique to her, engineers could swarm the data and quickly determine whether it affected many users. She says the resulting concrete examples create more empathy than dashboards alone and insists PMs must feel the customer’s pain.
A layered evaluation system governs Muse’s behavior
Mona says Muse has well over 100 evaluations, organized into several layers. Priority evaluations cover non-negotiables such as prompt injection, general safety, harm, and other security issues. Capability evaluations test functions such as browser use and task completion; behavioral evaluations measure intent understanding, follow-up questions, and the agent’s personality; and use-case evaluations cover priority activities such as shopping. PMs, engineers, researchers, and agents help write the evaluations, but humans must review the data and confirm that prompts represent product reality. The team even requires its system prompt and behavioral specification to be handwritten rather than generated by AI.
Word of mouth and starter ideas move activation
Mona says Muse’s strongest activation lever is word of mouth, not advertising. At a gathering of parents, a power user handed around her phone and showed how she used Muse; two people downloaded it immediately because they could relate to her examples. The second major lever is helping users send a meaningful first prompt. People often freeze and type “hi” or ask about the weather, so Muse includes an Ideas tab with packaged starter experiences that can create dashboards, scheduled notifications, or other useful outputs after the user says “let’s do it.” The team is now testing tiered onboarding because early adopters understand agents differently from later users who see Muse as just another chatbot.
Chat bubbles and one main chat were deliberate bets
Instead of copying the large-canvas interface used by many AI products, the team made Muse feel like a messaging service with chat bubbles. That choice made the agent feel more like an entity users communicate with, but created downstream design problems: long answers are awkward in bubbles, thoughts need to be split, and carousels can feel strange inside a message thread. The team also initially chose one main chat because Muse had persistent memory and long-running context, unlike assistants that required separate threads. Employee feedback split into “team main chat” and “team side chat,” so Muse eventually added side chats, although Mona remains “ride or die” for the main chat.
“Let them cook” defined the build culture
Muse’s small, engineering-led team encouraged engineers to take a kernel of an idea away for a few days and return with a working prototype. The internal phrase “let them cook” meant giving talented builders autonomy and trusting their product sense instead of forcing every idea through a formal specification process. One experiment put two users’ Muses, Fu and Quinn, in a group chat. The agents revealed a privacy problem when one disclosed a user’s doctor appointment to the other, then produced an entertaining but unsettling argument before their users made them negotiate peace. The team stopped the experiment after recognizing that discretion and multi-agent communication required much more work.
The first Muse prototype nearly failed at a board meeting
Alex Cornell produced the first working Muse prototype in only a few days, using a real model behind a simple interface. Mona initially found it strange, then realized she was using it differently from OpenClaw because it felt personal. The prototype was not supposed to appear at a board meeting: the pre-reads had already been distributed and Muse was barely an informal workstream. At the last minute, the team decided to show it, but the prototype broke about an hour before the meeting. Engineers and Alex repaired it five minutes before the meeting began, allowing the team to present it. Mona often imagines the alternate timeline in which they failed to fix it and Muse was never born.
What AI changes about product management
After more than 20 years as a PM, Mona says AI has changed prototyping, internal tooling, and the speed of product learning. A non-engineer can now build enough of an idea to make a concrete conversation possible, and teams can create internal tools that previously would never be prioritized. Because prototypes arrive quickly, teams debate what is actually in their hands rather than discussing abstract documents after months of engineering effort. The PM bar has therefore shifted toward product taste, customer empathy, and the ability to reject something that engineers enjoyed but users will not. Execution still matters, but Mona sees the PM’s central role as representing the customer and maintaining a high quality bar amid easy-to-produce AI slop.
A tiny team built Muse and now wants ambient agents
The original Muse group had roughly 10 people, mostly engineers, and the broader core team stayed under 100, with only about four PMs covering the product. The team intentionally kept decision-making and communication tight, sometimes letting engineers or designers lead early prototypes before assigning PM ownership for launch-ready areas. Mona says the experience was unusually close-knit: the team worked from a shared “library,” ate together, and built enough trust to move quickly. Looking ahead, she wants Muse to escape the app and become ambient in glasses, charm-like devices, and other hardware. She also wants agents to require less babysitting, become more financially useful, and help people start businesses, improve their health, save time, and achieve greater financial security.
Sponsors
- WorkOS · WorkOS provides drop-in APIs for enterprise features including SSO, SCIM, RBAC, and audit logs. The ad also describes AgentAuth for scoped, revocable agent identities, agent registration with AuthKit, and Airlock for allowing, denying, or escalating agent actions to humans. The host says companies including OpenAI, Anthropic, Cursor, Replit, Sierra, and Clay use WorkOS.
- DX · DX measures and benchmarks the impact of AI adoption across the product lifecycle. It shows where AI helps developers or creates bottlenecks, evaluates vendors, monitors costs and licenses, and identifies what limits engineering agents. The ad says companies including Snowflake, Sony, and BNY use DX.
Show notes
Muse is the #1 app in the App Store and arguably the fastest-growing AI consumer product of all time. Mona Sarantakos has led product on it since day one. She’s VP of Product at Meta Superintelligence Labs and a 10-year Meta veteran with over 20 years of experience as a PM. This is her first podcast interview.
We discuss:
1. How a weekend with OpenClaw on a Raspberry Pi turned into Muse
2. Why Muse waited seven months to launch, despite getting started before most of its competitors
3. How Nat Friedman, Alex Wang, and Mark Zuckerberg each left their imprint on the product
4. How the team uses “agent swarms” to surface user insights at scale
5. The use case nobody on the team predicted
6. How Jolly, a placeholder avatar, became an icon
7. The two levers that actually move activation for an AI agent
8. The #1 thing Mona looks for in PMs today
—
Brought to you by:
WorkOS—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more
DX—Engineering intelligence for the AI era
—
Episode transcript: https://www.lennysnewsletter.com/p/metas-muse-how-we-launched-the-most
—
Archive of all Lenny's Podcast transcripts: https://www.dropbox.com/scl/fo/yxi4s2w998p1gvtpu4193/AMdNPR8AOw0lMklwtnC0TrQ?rlkey=j06x0nipoti519e0xgm23zsn9&st=ahz0fj11&dl=0
—
Highlights: https://lennyspodcast.com/mostreplayedmoments
—
Where to find Mona Sarantakos:
• LinkedIn: https://www.linkedin.com/in/mona-sarantakos-a85873
• Threads: https://www.threads.com/@monachaudhuri
—
Where to find Lenny:
• Newsletter: https://www.lennysnewsletter.com
• X: https://twitter.com/lennysan
• LinkedIn: https://www.linkedin.com/in/lennyrachitsky/
—
In this episode, we cover:
(00:00) Teaser
(01:24) Introduction and Mona’s background
(02:14) A weekend with an early agentic AI
(09:08) The seven-month launch timeline
(14:07) "The claw is the law"
(15:53) Nat, Alex Wang, Zuckerberg, and how leadership shaped Muse
(23:37) The Jolly origin story
(29:53) Priority use cases and the shopping surprise
(35:03) Agentic commerce and how shopping works
(38:03) Agent swarms and synthetic data
(42:48) Evals: structure, hierarchy, and who writes them
(46:37) Activation and word of mouth
(51:09) The design bets that defined Muse
(57:04) Prototyping culture and the two Muses that fought
(01:01:54) How the PM role has changed
(01:09:05) Muse's first prototype
(01:10:37) The board meeting prototype that almost didn’t work
(01:12:00) Team size, closeness, and culture
(01:20:55) What comes next: ambient agents and financial betterment
—
Referenced:
• Muse: https://muse.ai
• Introducing Muse Spark: MSL’s First Model, Purpose-Built to Prioritize People: https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs
• Nat Friedman’s website: https://nat.org
• OpenClaw: https://openclaw.ai
• From skeptic to true believer: How OpenClaw changed my life | Claire Vo: https://www.lennysnewsletter.com/p/how-openclaw-changed-my-life-claire-vo
• Personal Superintelligence: https://www.meta.com/superintelligence
• The Future is for Everyone: https://www.meta.com/thefutureisforeveryone
• Grok Bot: https://x.ai/bot
• How we built Grok Bot in a month | Roman Ugarte (SpaceXAI): https://www.lennysnewsletter.com/p/how-we-built-grok-bot-in-a-month
• Claude Duet: https://claude.com/product/claude-duet
• Dots: https://openai.com/index/introducing-dots
• Alex Wang on X: https://x.com/alexandr_wang
• Alex Cornell’s website: https://www.alexcornell.com
• How I AI: How Meta designed Muse with Alex Cornell: https://www.lennysnewsletter.com/p/how-i-ai-how-meta-designed-muse-with
• Christine Awad on X: https://x.com/christinesed
• Linear’s secret to building beloved B2B products | Nan Yu: https://www.lennysnewsletter.com/p/linears-secret-to-building-beloved-b2b-products-nan-yu
—
Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [email protected].
—
Lenny may be an investor in the companies discussed.
To hear more, visit www.lennysnewsletter.com
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account