Source: raw/Meta_s_Muse_-_Inside_the_launch_of_the_most_successful_consumer_product_this_year_Mona_Sarantakos.md
Creator: Lenny’s Podcast (Lenny Rachitsky; guest: Mona Sarantakos, head of product for Muse and VP of product at Meta Superintelligence Labs) · URL: https://www.youtube.com/watch?v=WY0cdD4gNoQ · Duration: not stated in the transcript (about 18,000 words) · Platform: YouTube (auto-captions, fetched 2026-11-17)
A first-party account of how Meta built its consumer personal agent, Muse, from the person who ran product. It covers why the team took about seven months when rivals took two, how evals were organised, which design bets were contested inside the company, and what does and does not get new users to their first useful task. Sarantakos gives no usage or revenue numbers. The transferable material is the eval hierarchy, the onboarding findings and the rules about what must be written by hand.
Key Takeaways
- Timeline. Nat Friedman pinged a few people at the end of January or start of February after trying OpenClaw; by February the group was sure it wanted to pursue it; by spring there was a working end-to-end product; alpha testers and employees used it through the summer. Lenny puts idea-to-launch at seven months “from what I’ve read”, against about two for Grok Bot, Duet and Dots.
- Why the wait: “trust is a must.” The time went to reliability, safety and security. “If we launch too early and we have a bunch of issues, people will not trust this agent. If you don’t trust the agent, you’re never going to use it.” She would not change the decision and says there has been no “catastrophic problem” so far.
- Industry benchmarks were not enough. For an agent with personal context that can buy things, the team built its own bar and spent “a lot of team time red teaming”. Each new feature reopened the work: adding a browser brought “a whole new set of problems” for security.
- A backstop outside the agent. She describes a “sentinel”, designed with Friedman, the architects and the security team, that sits outside the core Muse agent as a secure backstop.
- Team size. About 10 people at first, mostly engineers plus designer Alex Cornell. Under 100 “all in”, with infrastructure and security teams alongside. Four PMs built the whole thing, so many areas had no PM.
- No formal reviews. Leadership feedback came from handing a phone with a prototype to Mark Zuckerberg, Friedman or Alex Wang in small rooms. She made her first slides in August, shortly before launch.
- Model surprise. The first Muse Spark model “was way more agentic than we thought it was going to be”. She credits Wang with pushing the launch: “guys, I think it’s ready.”
- Shopping was the surprise use case. She had predicted personal productivity, health and personal finance; others predicted home automation. Dogfooding showed shopping, which is now “one of our top use cases”. Home automation faded once the user base widened beyond early adopters.
How the evals are organised
- Four families:
- safety (prompt injection, general safety, harm);
- capability (browser use, task completion);
- behavioural (personality, follow-up questions, intent understanding);
- use-case evals for the priority use cases.
- Three tiers of importance: “always priority” evals that every change must pass (safety is the example); secondary evals where a small regression can be traded; feature-level evals owned by specific teams.
- Evals follow usage. When shopping emerged, “all of our evals have been over here”, so the team built shopping evals.
- Who writes them: PMs some, engineers and researchers many, with agents helping. People must review the data: “when we don’t, we always have issues.”
- What must be handwritten: the system prompt (“you cannot use an AI to write a system prompt, any part of it”), the behavioural spec, and a human check that the eval prompt distribution matches real product use.
- The data problem. Meta AI conversations were mostly single-turn. Muse is multi-turn over long-running, non-coding tasks, and no production data like that existed.
- Agent swarms filled the gap. Agents “swarm out” over conversations and usage data, produce synthetic data, and return formatted “dumps” of what users attempt and where they are blocked. The team now swarms on each new model and to check whether a problem one person spotted is real. Her example of the output: ten people tried to shop on a site and got blocked. She has every PM read them, because they bring back insights, not only numbers.
Activation: what gets a new user going
- Her number one problem: “I already have a chatbot. Why do I need another one?”
- Word of mouth beats advertising. At a parents’ gathering, a power user passed her phone around and two people downloaded Muse on the spot. No ad line (“personal superintelligence”, “an AI that does things for you”) landed until they saw someone like them using it.
- People freeze on the first prompt. The usual first message is “hi” or the weather. The Ideas tab offers starter packs where the user only has to say “let’s do it” and Muse builds a scheduled notification, a dashboard, or does the task.
- Onboarding must be tiered. What worked for early adopters, who already knew what an agent was, does not work for later cohorts asking how this differs from other AI products. Meta now serves onboarding variants by where a user comes from and their likely exposure.
- Not solved. “I don’t think we figured it out.”
Design decisions that were contested
- The avatar, Jolly. It began as a one-shot placeholder in Cornell’s prototype (“Jollybot”). People kept saying it would be replaced by a logo; it was nearly cut, and kept about a week before launch with Zuckerberg’s support. Her reason: “when you’re talking to something that feels so personal, the last thing you want to talk to is a corporate logo.” Working animations (headphones, a laptop) came from the team being tired of an avatar that stared at them.
- The principle behind it: the Muse belongs to the user, and Meta “needs to fade away”, within guardrails, a charter and the law. Stating that early made later arguments short.
- Chat bubbles, not a canvas. Other AI products, Meta AI included, were moving to a full-width essay layout. Muse copied a messaging app because that is how people talk to an assistant. The cost is unresolved questions about splitting long answers and whether a carousel belongs in a chat thread.
- One main chat. With long-running memory, the team shipped a single thread. Employees split into “team main chat” and “team side chat”, and side chats were added. She still uses only the main chat.
- Goals over tasks. Zuckerberg reframed the purpose as helping people reach their goals, not completing tasks. That led to “dreaming and study” (the agent thinking about the user’s priorities between sessions) and the Goals tab.
Stories with a lesson
- Discretion, learned the hard way. Two employees put their Muses in a group chat. One Muse told the other’s owner about a doctor’s appointment. The two agents later argued, passive-aggressively, until a human intervened. The team concluded multi-agent chat was not ready.
- First refund. On a trip to the World Cup, two cancelled flights were refunded after the team members’ Muses dealt with the airline’s agent.
- Shopping behaviour changes. She used to bulk-order from one store; Muse now buys from five to seven stores to match her spec, which means more parcels. She does not expect browsing to disappear and describes retail partners testing business agents that talk to consumer agents.
- The board demo. The prototype broke an hour before a board meeting where it was not on the agenda and was fixed five minutes before.
What she says changed for PMs
- Prototypes replace documents. Discussion is about something in your hand, so sunk cost is lower and decisions move faster. She can test an idea within a day, sometimes herself.
- Taste is the hiring bar. With engineers who have product sense, the PM’s job is to be “the voice of the customer” and to hold the quality bar, because “AI slop is very easy to put out into the world”.
- “Let them cook.” Many features came from one or two engineers going away for a few days with the start of an idea.
- PMs come in later. Zero-to-one work often needed a builder, not a PM; a PM joined once a feature was headed for launch.
Try It
- Sort your evals into three tiers: must-pass, tradeable, feature-level. Decide which tier each one is in before the next model swap.
- Write the system prompt and behaviour spec by hand, and have a person check that your eval prompts look like what users really send.
- Add a “let’s do it” starter to any agent or assistant you ship: a pre-packaged first task the user accepts with one tap, instead of an empty prompt box.
- Collect and show real user stories before spending on positioning copy. Her evidence is that a peer’s phone converts and a tagline does not.
- Split onboarding by audience: people who already use agents need the difference explained; people who do not need the category explained.
- Before letting two agents share a channel, decide what each may disclose about its owner.
Open Questions
- No metrics. Lenny says Zuckerberg “has hinted” Muse is the fastest-growing AI consumer product and calls it number one in the App Store. Sarantakos confirms neither with a figure.
- Team size against the earlier figure. She says under 100 for the Muse team. Wang’s “fewer than 200 … across the model and the product”, recorded in the Nate B Jones article, covers a wider group, so the two may both hold.
- Names from auto-captions. OpenClaw is captioned as “Oculot”, “Open Clau”, “Open Clot” and “Open Claw”. The head of engineering is captioned “Christina Wod” and then “Christine”, and the tech lead as “Tarek”; these spellings are unverified. Wang’s first name is given only as “Alex”.
- The sentinel is mentioned once and not described further.
- Future hardware. She says Muse should move to glasses and “charm devices”; Lenny refers to a “charm” device launching at some point. No date or detail is given.
Related
- Muse and the Case for Another Swing — an outside reading of the same launch, with Wang’s team-size and behaviour-testing figures
- Personal Agents Compared (October 2026) — Muse beside Grok Bot, Dots and Duet
- Claire Vo’s Personal Agent Benchmark — how Muse performed on nine tasks
- Dianne Penn — Evals Are the New PRDs — another product lead on evals as the core PM artefact
- Tulsee Doshi on Gemini PM Evals — a second eval-organisation scheme to compare
- AI Shopping and Agentic Commerce — what agents buying on a user’s behalf means for merchants
- What Personal Agents Keep — the access and discretion questions her group-chat story raises