Mona Sarantakos: How Meta built trust into Muse before launch

Muse’s product lead explains its seven-month build, the surprise popularity of shopping, and why real user stories are its strongest activation tool.

MSMona SarantakosGuest · VP of Product
Channel
Lenny's Podcast
Published
Nov 15, 2026
Video
1:27:15
Read
12 min
Original
English
Open on YouTube
Scroll to read and the video shrinks to the bottom-right corner. Captions follow your reading language.
Reading languageAvailable in 2 languages
You're reading the original English.
SUMMARY

In 30 seconds

Meta’s Muse team spent months on reliability, safety and security because people would not give a personal agent meaningful work without trust, says product lead Mona Sarantakos. Early users also overturned its assumptions: shopping grew into a leading use case while home automation mattered less than expected. Sarantakos describes using agent swarms and human-reviewed evaluations to understand long-running tasks, and says relatable stories plus guided first actions help people grasp what an agent can do.

  1. 1Safety and reliability took months because a personal agent must earn trustSarantakos says a working prototype came relatively early, but browsing and actions such as purchases created security and evaluation work before launch.
  2. 2A temporary mascot survived because the agent should feel personalDesigner Alex Cornell’s placeholder Jolly led to animations and customizable avatars; the team felt a corporate logo would work against the product’s ownership principle.
  3. 3Shopping became a leading use case the team did not predictFeedback from internal and external users prompted the team to add shopping evaluations; home automation was less prominent than expected.
  4. 4Agent swarms helped build evaluations for long, non-coding tasksSarantakos says the team used feedback, examples and synthetic data to understand multistep conversations, with people reviewing the resulting evaluation data.
  5. 5Relatable examples and an easy first action help people activateShe says word of mouth is the most effective lever and describes an Ideas tab that starts a task for users who do not know what to ask.
  6. 6Bots sharing a conversation exposed a discretion failureDuring an early experiment, one bot disclosed another person’s private appointment to the group, highlighting a problem for shared-agent experiences.
  7. 7Prototypes let product teams debate something concreteSarantakos says faster prototyping changes product work, while representing user needs and judging what is good enough remain central.

Takeaways

  1. A personal agent needs a high reliability, safety and security bar before people will entrust it with sensitive context or actions, Sarantakos argues.
  2. Track actual use cases: shopping surprised the Muse team and caused it to create dedicated evaluations.
  3. Combine AI-assisted analysis with human review of evaluation data; the team handwrites its system prompt and behavioral specification.
  4. Show people how someone like them uses the product, then make the first useful action easy to start.
  5. Rapid prototypes help product teams test ideas concretely before committing months of engineering work.
  6. Agent-to-agent collaboration requires discretion: a test bot shared another person’s private appointment with the group.
FULL TEXT

Read the edited text by chapter

Transcript only
01

A weekend with OpenClaw changed the product brief

Sarantakos saw the potential of a personal agent after connecting an early tool to her own information. A small group began experimenting before Muse became a formal project.

Before Muse, Mona Sarantakos was working on Meta AI. Her team was reconsidering the assistant from first principles while anticipating the first Muse Spark model. The question was whether to keep iterating on what existed or start again.

Around the end of January, Nat Friedman urged a few colleagues to try OpenClaw. Sarantakos had imagined consumer agents becoming important toward the end of 2026 or in 2027, but his message pushed her to experiment immediately.

Unable to find a Mac mini, she set OpenClaw up on a Raspberry Pi and spent a weekend talking with it. She connected email, calendar and home automation, taking a much more permissive approach than her husband considered safe.

On Monday it assembled a briefing from school newsletters and other sources. Keeping track of the children’s activities had been difficult; with only a small prompt, she felt unexpectedly prepared for the week. That experience made her think about other busy parents and families.

She told Friedman she wanted to pursue the idea, even though there was not yet a formal product or team to join. A few colleagues stepped away from their existing work to explore what a personal agent could become.

Their experiments also showed why a consumer product would need to differ from the tool they were trying. Sarantakos lists easier setup, greater safety and scalability as problems they would have to solve, so the group prototyped before committing to an official project.

02

Why a working prototype was not ready to launch

The team spent months on reliability, safety, security and testing. Sarantakos describes a more direct, prototype-led relationship with company leaders.

Sarantakos places the first conversation around late January and says the team became confident enough to pursue the project seriously in February. Lenny Rachitsky asks why, despite starting early, the product took about seven months to launch.

The work had several phases. First the team had to decide what the product actually was, solve infrastructure questions such as running virtual machines, and test interface choices. A functioning end-to-end version existed by spring, she says.

Reliability, safety and security took much longer. Existing industry benchmarks did not feel sufficient for an agent with personal context that could browse and make purchases. The team had to define its own evaluation thresholds and spend substantial time red-teaming.

Adding a browser created new security questions. Sarantakos says the team wanted to pass quality gates rather than release before users could reasonably trust the product; an early failure, in her view, could discourage people from ever relying on it.

She would not change the decision to take that time. She says the team gradually expanded employee use and external alpha testing through the summer, so it was receiving user feedback even before a broad launch.

The internal phrase “the claw is the law” came from the team’s effort to explain the promise it saw in OpenClaw to colleagues who had not experienced it and often found the idea alarming. The phrase became a shorthand for the direction they wanted to explore.

Sarantakos describes the Muse group as small and engineering-led, with an unusually direct relationship to leadership. Instead of formal presentation cycles, the team often brought working prototypes into small meetings and passed a phone around for feedback.

Those conversations could run much longer than scheduled. She says the product emerged from a series of informal discussions with Nat Friedman, Alex Wang and Mark Zuckerberg, sometimes together and sometimes individually.

Friedman was involved in detailed architecture and security choices, including an external backstop the team called a sentinel. Sarantakos describes him as closely involved with the engineers rather than merely approving finished plans.

She says the initial Muse Spark model was more capable at agentic work than expected. Wang helped align research priorities with product needs and, in her account, eventually encouraged the team to launch.

Zuckerberg pushed the team to frame the product around helping people reach their goals, rather than just completing isolated tasks. That framing led to a goals tab and to ideas for an agent that can use what it knows to make more useful suggestions over time.

Sarantakos calls some of that work “dreaming and study”: the possibility that an always-available agent can consider a user’s priorities between requests. She presents it as the team’s product direction, not a guarantee that the agent understands every goal.

03

Jolly survived as the product became personal

A temporary avatar inspired animations and customization. The team defended it as consistent with an agent that belongs to its user.

Designer Alex Cornell created Jolly as a temporary creature for an early prototype. It originally slept and woke when the app opened; the name came from a nickname for his daughter. The team expected that a corporate logo might eventually replace it.

Once the character was in use, the team wanted it to move and show what it was doing—for example, by wearing headphones while working. That led to animations and then to the idea that each user should be able to create a different avatar.

Sarantakos says there was a real point at which an alternative logo might have replaced Jolly. The team and Zuckerberg preferred the avatar: in their view, a corporate mark felt at odds with a personal agent that is meant to work for its user.

The decision also illustrated how product teams can exercise judgment inside a large company. Sarantakos says stating the underlying principle—Muse belongs to its owner, within legal and product guardrails—made the design choice easier to defend.

04

Actual use changed the team’s priorities

Shopping became more important than expected, while home automation was less prominent. The team gathered early feedback and built use-case evaluations.

Rachitsky asks how the team handled unusual use cases from senior leaders, such as Zuckerberg wanting feedback on martial arts. Sarantakos says the group spent hours discussing use cases but tried to ground decisions in stories from actual users.

An external research panel started in March, while employees also used an internal feedback group. The team brought examples of what did and did not feel useful to leadership roughly weekly, rather than treating one executive’s preference as the whole market.

Shopping surprised Sarantakos. She had prioritized productivity, health and personal finance, while some colleagues expected home automation to be large. As shopping emerged in testing, the team built dedicated evaluations; home automation mattered less for the broader audience than expected.

She describes engineering manager Christine Awad as an enthusiastic shopping tester who regularly sent bug reports. Sarantakos found that asking an agent to buy items from several stores could produce better variety, but also many separate deliveries.

On agentic commerce, she does not think buying should eliminate browsing. People sometimes just need a routine item, but at other times enjoy looking for inspiration. Her hope is that agents narrow choices and make trade-offs without removing the enjoyable part of shopping.

She says retailers and merchants are experimenting too. In a possible future, consumer and business agents could exchange useful information, but the interaction between discovery, brands and purchase remains an area the industry is still working out.

05

Turn long-running user experiences into evaluations

Agent swarms helped examine feedback and generate data, and evaluation sets cover safety, capabilities, behavior and use cases. People still review data and write key instructions.

The team also used what it calls agent swarms to gather and organize signals from feedback, data and examples, and to help create synthetic data. These summaries helped the team understand which tasks worked and where users were getting stuck.

A challenge was that much prior assistant data was single-turn, whereas Muse conversations involved extended back-and-forth over non-coding tasks. Sarantakos says the team needed examples of long-running work such as meal planning to construct useful evaluation sets.

When she suspects a problem but cannot tell whether it is personal to her experience, engineers can investigate with a swarm. She says this lets the team move from a dashboard trend to specific examples, such as repeated failures on a particular shopping site.

Sarantakos asks product managers to look at these concrete reports because they can make users’ difficulties feel real. For her, understanding that pain is part of deciding which problems deserve attention.

The team groups evaluations into safety, capabilities, behavior and priority use cases. Safety includes prompt-injection and harm tests; capability checks include browsing and task completion; behavioral tests examine intent understanding and follow-up questions.

Not every evaluation has the same status. Some priority tests must pass for a change to proceed, while other regressions may involve trade-offs. Individual features also have their own more specific tests.

Product managers, engineers and researchers all write evaluations, sometimes with agent help. Sarantakos stresses that people must review the data and check whether prompts reflect the product’s actual distribution of use.

The team has recently required its system prompt and behavioral specification to be handwritten. Sarantakos says the rule reflects past problems when important data was not reviewed; the organization still uses agents heavily for other parts of the process.

06

Show a first useful action and design for conversation

Sarantakos sees word of mouth and guided starter ideas as ways to help new users activate. She explains the trade-offs behind chat bubbles, one main chat and later side chats.

Rachitsky asks how people reach an initial useful moment with an assistant. Sarantakos says many people already have a chatbot and do not understand why they need another product or what an agent can actually do.

In her experience, word of mouth is the strongest activation lever. At a gathering of parents, one experienced user passed her phone around to show how she used Muse; Sarantakos says two people downloaded it after seeing examples they could relate to.

The second challenge is getting people past the first prompt. Someone may start with a greeting or weather question because they do not know the range of possibilities, even if the underlying system can perform more involved tasks.

The Ideas tab offers starter packs for actions such as setting a recurring notification or building a dashboard. A user can accept a suggestion and see a task begin, instead of first having to invent a detailed instruction.

Onboarding also differs by audience. Sarantakos says early adopters often arrived knowing what an agent was and wanted to test safety or capability, whereas later users may first need to understand how the product differs from familiar AI chat.

The team is testing different onboarding paths based on prior exposure, but she does not claim it has solved the problem. As more people use agents and share examples, she expects understanding to improve.

Another unusual design choice was to use chat bubbles rather than a broad, essay-like canvas. The team wanted communication to feel familiar, like a messaging service with a personal assistant.

That choice created new constraints: long responses need breaking into readable messages, and the team must decide which richer interface elements fit naturally inside chat. Sarantakos says some of those questions are still unresolved.

The team initially favored one main conversation because a persistent agent could remember context across turns. Employee testing was polarized: some wanted separate threads for mental organization, while others preferred a single continuing conversation.

Muse eventually added side chats in response to demand, while the agent remains aware of conversations across them. Sarantakos still prefers the main chat and says the team expects to continue refining the design.

She says the main-versus-side-chat decision was made by the product leads. Some choices clearly needed senior leadership, but the team had latitude over many others and leaders often asked what the team itself wanted to do.

07

Prototypes, discretion and changing product management

Engineers had room to prototype, but a bot-to-bot test exposed a privacy failure. Faster prototypes also let product managers test ideas concretely while retaining responsibility for product judgment.

“Let them cook” described the team’s prototyping culture. A small number of engineers could take an initial idea away for a few days and return with a working version that showed possibilities even the person who suggested it had not anticipated.

In an early experiment, two users placed their agents in a shared chat. One agent mentioned that its owner had visited a doctor, information the other participant did not know. The incident exposed how difficult discretion becomes when agents carry personal context into shared spaces.

The agents also argued with each other, and their owners eventually intervened. The experiment was entertaining to watch, she says, but the team concluded there were substantial problems to solve before this kind of shared interaction was ready.

Sarantakos contrasts that experimentation with an older product process that required a fully formed specification and a prediction of every edge case. With prototypes arriving faster, she can change her mind based on what she sees rather than wait for a lengthy build.

For product managers, she says the ability to create a rough prototype or internal tool without being an engineer changes the discussion. A concrete object gives colleagues something to react to instead of relying on a document that everyone imagines differently.

A prototype also reduces the cost of discovering that an idea is wrong. Previously, once engineers had spent months building a concept, the team could feel reluctant to reject it; faster experiments make that conversation easier.

Execution still matters, but Sarantakos says product judgment and representing customers have become especially important. When software can be built and deployed quickly, a product manager needs to decide whether the result is actually useful and meets a high enough bar.

Rachitsky asks whether design was deliberately chosen as a competitive differentiator. Sarantakos says Cornell’s design was intentional, but the team did not begin with a plan to differentiate through design; seeing his work helped them articulate their product principles.

08

A small team built around working prototypes

Cornell’s early prototype nearly failed before a board meeting, and a small cross-functional group grew around it. Sarantakos describes the benefits and strain of a lean team.

Cornell created the first working Muse prototype in days, with a chat and feed. At first Sarantakos found it strange, but using it on his phone made the personal nature of the interaction clearer to her.

The prototype was not in the planned board-meeting materials, but the team decided to show it at the last minute. It broke about an hour beforehand and was repaired roughly five minutes before the meeting, she recalls.

Among early moments, she remembers two agents communicating with airlines and obtaining refunds after flights were disrupted. The experience showed the team a task it had not expected the agents to complete.

Her strongest memory, however, is the team becoming close. A shared space in a Menlo Park office became the place for standups, meals and celebrations as the group spent long stretches building together.

Sarantakos estimates the first team at about ten people, mostly engineers and the designer. It later grew to fewer than 100 across functions, with outside infrastructure and security teams also involved; she says the core remained intentionally small.

There were about four product managers. The team had to decide where a PM was essential and where an engineer or designer could lead initial exploration. A PM became more important when a prototype was going to be integrated or prepared for launch.

A small team kept communication and decisions close, but it also meant little slack and lost sleep. As the team grows after launch, Sarantakos says preserving the early culture is a new challenge.

She calls Muse unusually personal because the team first built it for themselves. As more people use it for tasks involving family life, schedules and health, she also feels greater responsibility for what can go wrong.

The nondeterministic nature of agents makes product management different from a fixed flow: guardrails and evaluations help, but the agent still determines how to act in a particular context. Different users’ agents can behave differently because their contexts differ.

09

The ambition and uncertainty after launch

Sarantakos feels responsibility for users who rely on the agent for personal tasks. She hopes for more capable agents across devices, while acknowledging that the path remains uncertain.

Looking ahead, Sarantakos hopes an agent will be available beyond an app, potentially through glasses or other devices. She also wants it to need less supervision: if caring for the agent takes as much effort as doing a task, it has failed to save time.

She is interested in ways agents might help with financial security or health as well as saving time. These are ambitions, and she says she cannot predict the precise path between today’s product and an agent available everywhere.

Asked how listeners can help, Sarantakos asks for feedback through the product. She says those reports reach the team and are useful because the product and the broader agent field are still early.

More from this channel

58:33EN
Lenny's Podcast· Nov 8, 2026Podcast · 2 speakersGemini’s product chief on testing models and building ahead of themGuest Tulsee Doshi · Host Lenny RachitskyTulsee Doshi explains a small-eval starting point, her 60-20-20 product framework and why model behavior itself is a user experience.Product ManagementFoundation ModelsEvals+2
1:27:47EN
Lenny's Podcast· Nov 1, 2026Podcast · 2 speakersHector Serrano on Building a Paid Creator Program at Wispr FlowGuest Hector Serrano · Host Lenny RachitskyHow a creator-led growth team recruits, compensates and supports paid UGC creators—and uses founder storytelling alongside ads.Growth StrategyCreator EconomyContent Strategy+1
1:25:47EN
Lenny's Podcast· Oct 25, 2026Podcast · 2 speakersRob Snyder’s PULL Framework for Finding Early DemandGuest Rob Snyder · Host Lenny RachitskyA practical conversation about customer urgency, alternatives, early sales and retention.Company BuildingProduct Strategy
1:20EN
Lenny's Podcast· Sep 4, 2026Podcast · 2 speakersTara Seshan on OpenAI’s Vision for ChatGPT and WorkGuest Tara SeshanA product lead explains a goal of simplifying model and interface choices.OpenAIProduct & UXAI agents
0:46EN
Lenny's Podcast· Dec 31, 2025PodcastA View on Effort and Team PerformanceA short argument that difficult moments can reveal how teams respond under pressure.LeadershipCompany Culture
0:23EN
Lenny's Podcast· Aug 31, 2023PodcastWhy Product Quality Matters After the SaleGuest Meltem Kuran BerkowitzMeltem Kuran Berkowitz on customer expectations in business software.Product StrategyProduct Management