Meta's personal AI agent app with a single main chat and a mascot called Jolly.
Meta's Muse, a personal AI agent with one main chat and a mascot called Jolly, launched after seven months of reliability and safety work, built by a team under 100. It became a top consumer AI agent with no catastrophic problems so far. Lesson: hold agents touching personal data to a self-built eval bar until trust is earned, then lean on word of mouth for activation.
From Mona Sarantakos on building Muse
Muse started as an informal side experiment, not a planned product. Nat Friedman pinged a few colleagues at the end of January, and Mona Sarantakos ran OpenClaw on a Raspberry Pi because Mac Minis were sold out. A few-word prompt produced a Monday briefing of school and calendar items that made her feel on top of her personal life for the first time.
She was rebuilding Meta AI from scratch at the time, and her H1 brief had guessed consumer agents would arrive around end of 2026 or 2027. She connected email, calendar and home automation; her husband said, "That is so not safe." The team stayed informal while prototyping and mobilized a proper team only once they were sure there was a product.
A short vision document gave a scary idea a name to rally around. "The Claw Is the Law" was written to explain why the team believed in this kind of agent, since most people had never heard of OpenClaw and it sounded unsafe. It became a catchphrase and a north star.
Mona says it hasn't been used in a while.
Each senior leader shaped a different layer of the product. Nat Friedman shaped the engineering, including the Sentinel, a backstop outside the core Muse agent. Alex Wang adjusted research priorities after the first Muse Spark model proved far more agentic than expected, and pushed the launch: "I think it's ready. Just go."
Work with leadership ran like a startup: Mona made no slides until August, just before launch, and meetings booked for minutes with Zuckerberg could run an hour. After launch Wang became "our best marketing machine ever" on social media.
Zuckerberg moved the goal from task completion to helping users reach goals. The team had fixated on tasks such as making purchases; Zuckerberg pushed back to "the purpose of Muse is to make people's lives better", which requires knowing the user's priorities. That produced the explicit goals tab and the idea of "dreaming", a 24/7 agent that thinks about the user and makes proactive suggestions.
He stressed alignment: knowing what the user is trying to do. Engineers' eyes lit up once proactive help became the goal.
Muse held launch until reliability, safety and security met its bar. A working end-to-end product existed by spring; most of the seven months went to custom evals, red teaming, employee dogfooding and external alpha tests with growing groups. Muse launched later than rival agents and has had no catastrophic problem so far.
The internal saying is "trust is a must", because "if you don't trust the agent, you're never going to use it." Mona argues an early launch with a bunch of issues loses trust for good, and says she would change nothing about the timing.
The team wrote its own eval bar because industry benchmarks didn't fit. No benchmark covered an agent holding personal context that can buy things, so Muse set its own bar. Every new capability reopens the security surface: adding a browser created new attack paths that had to be closed before shipping.
Quote: "the benchmarks that are out in the industry just didn't feel sufficient".
Evals are tiered: must-pass, tradeable secondary, then feature level. Evals are grouped into safety (prompt injection, general harm), capability (browser use, task completion), behavioral (follow-up questions, intent understanding) and use case. Priority evals like safety must always pass, secondary ones can be traded off slightly, and teams add feature-level evals.
PMs, engineers and researchers all write evals, with agents helping. Muse has too many to count.
Humans handwrite the system prompt and behavioral spec. A recent Muse rule says no part of the system prompt may be AI-written. People also review eval data to confirm its prompt distribution matches real product use.
Mona says issues appear whenever the team skips that review.
Agent swarms generate multi-turn data and find real production issues. Meta AI conversations were mostly single turn, so there was no data for long-running multi-turn evals. Agents swarm over conversations and synthetic data and return formatted dumps of what tasks people did and what worked or failed; the team swarms every new model and any suspected issue.
One suspected problem was swarmed, confirmed as real and given an owner. Findings "feed into our evals".
PMs read specific production examples, not just dashboards. Swarm output shows concrete cases, such as 10 people asking to shop on a site and getting blocked, which lets the team target fixes precisely. Mona makes all her PMs do this to build empathy.
"if you don't have that empathy and you can't feel the pain, you will never develop good products."
Agents in a group chat taught the team about discretion. Early on, two employees' agents shared a group chat and one told the other person its owner had been to the doctor. The agents later had a passive-aggressive fight, and the team decided multi-agent chat was not ready to ship.
"we learned about discretion the hard way"
A throwaway mascot beat the corporate logo because Muse belongs to the user. Lead designer Alex Cornell made Jolly with a one-shot prompt, named after his nickname for his daughter. A replacement logo was ready, but about a week before launch Zuckerberg's love of the avatar, team resistance and employee affection kept Jolly, and users can now make their own avatars.
Mona calls Jolly "our cockroach" because every few months someone said they couldn't launch with a stuffed animal at the top. Added animations like headphones and a laptop built more system around it. "When you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo."
An early principle settled design fights instead of opinion. The team stated early that the agent works for and is owned by the user, with guardrails and a charter from Meta. Once that existed, a corporate logo clearly violated it.
Mona tells PMs they have agency to push back on defaults: "it's not a random rationale".
Chat bubbles make the agent feel like messaging a person. While most AI products moved to large canvas, essay-style answers, Muse chose chat bubbles. That constrains long answers and forces choices like how to break up thoughts and whether a carousel feels odd.
The choice was controversial inside Meta and the team is still working through the downstream effects.
One main chat shipped first; side chats came only after demand. The team argued an assistant with long-running context should be one conversation, like a personal assistant who remembers everything. Employee reaction was the most polarizing of the first release, and side chats were added once demand for mental separation was clear.
Mona still uses only the main chat.
Word of mouth activates users that advertising could not. Mona watched a power-user mom pass her phone around a group of parents and scroll through how she used Muse; two of them downloaded it on the spot. The team tries to make it easy for users to share real stories and examples.
Activation is now the hardest problem. "the most effective lever out there is word of mouth."
Starter packs in an ideas tab fix first-prompt freeze. Most first prompts are "hi" or "what's the weather". The ideas tab offers packs, such as creating a notification cron or building a dashboard, where the user just says "let's do it" and Muse starts work.
Mona says it has gone relatively well, but people still struggle to imagine what the agent can do.
Later adopters get different onboarding than early adopters. Early adopters knew agents and wanted Muse to be safer and push boundaries; later cohorts ask how it differs from AI they already use. Muse now serves onboarding variants based on where users come from and how exposed they are to agents.
The team says it has not figured onboarding out yet.
Shopping became a top use case nobody predicted. Mona expected productivity, health and personal finance, others expected home automation, but dogfooders started shopping. Because evals existed per use case, the team refocused and added shopping evals.
Home automation, big in the OpenClaw days, did not take off with the broader base. One engineer shopped so much Mona joked about taking her credit card away.
From Mona Sarantakos on building Muse
Under 100 people and four PMs built the whole product. The first team was about 10 people, mostly engineers plus designer Alex Cornell, and the effort stayed under 100 including all functions. A PM joined an area once it was headed for the product or launch.
"communication management and decision making had to be very very tight"
Engineers prototype from a kernel of an idea before specs. Many features began with one or two engineers taking an idea away for a few days and returning with a prototype. Cornell built the first Muse prototype, a chat thread and a feed, in days; it was shown to the board after being fixed minutes before the meeting.
Mona says this needs a talent-dense team with product sense and a lot of trust: "You got to let the engineers cook."
Reviewing working prototypes replaces slide reviews and avoids sunk cost. Leaders reviewed phone prototypes in small rooms, and the team could say "I hate this" without months of engineering behind it. Mona can now get to an idea in a 24-hour cycle, sometimes building it herself.
She also uses AI to build internal tools that never got prioritized.
Product taste is the top quality Mona hires PMs for. When building is easy, shipping AI slop is easy too, so PMs earn their place as the voice of the customer holding a high bar. Execution still matters, but taste comes first.
"product taste now is the most important thing."
Launched after seven months of reliability, safety and security work, built by a team under 100 with about four PMs. It became a top consumer AI agent, and the hardest problem now is activation, where word of mouth works best. Shopping surprised the team as a top use case.
Muse started as an informal side experiment, not a planned product. Nat Friedman pinged a few colleagues at the end of January, and Mona Sarantakos ran OpenClaw on a Raspberry Pi because Mac Minis were sold out. A few-word prompt produced a Monday briefing of school and calendar items that made her feel on top of her personal life for the first time.
She was rebuilding Meta AI from scratch at the time, and her H1 brief had guessed consumer agents would arrive around end of 2026 or 2027. She connected email, calendar and home automation; her husband said, "That is so not safe." The team stayed informal while prototyping and mobilized a proper team only once they were sure there was a product.
A short vision document gave a scary idea a name to rally around. "The Claw Is the Law" was written to explain why the team believed in this kind of agent, since most people had never heard of OpenClaw and it sounded unsafe. It became a catchphrase and a north star.
Mona says it hasn't been used in a while.
Each senior leader shaped a different layer of the product. Nat Friedman shaped the engineering, including the Sentinel, a backstop outside the core Muse agent. Alex Wang adjusted research priorities after the first Muse Spark model proved far more agentic than expected, and pushed the launch: "I think it's ready. Just go."
Work with leadership ran like a startup: Mona made no slides until August, just before launch, and meetings booked for minutes with Zuckerberg could run an hour. After launch Wang became "our best marketing machine ever" on social media.
Zuckerberg moved the goal from task completion to helping users reach goals. The team had fixated on tasks such as making purchases; Zuckerberg pushed back to "the purpose of Muse is to make people's lives better", which requires knowing the user's priorities. That produced the explicit goals tab and the idea of "dreaming", a 24/7 agent that thinks about the user and makes proactive suggestions.
He stressed alignment: knowing what the user is trying to do. Engineers' eyes lit up once proactive help became the goal.
Muse held launch until reliability, safety and security met its bar. A working end-to-end product existed by spring; most of the seven months went to custom evals, red teaming, employee dogfooding and external alpha tests with growing groups. Muse launched later than rival agents and has had no catastrophic problem so far.
The internal saying is "trust is a must", because "if you don't trust the agent, you're never going to use it." Mona argues an early launch with a bunch of issues loses trust for good, and says she would change nothing about the timing.
The team wrote its own eval bar because industry benchmarks didn't fit. No benchmark covered an agent holding personal context that can buy things, so Muse set its own bar. Every new capability reopens the security surface: adding a browser created new attack paths that had to be closed before shipping.
Quote: "the benchmarks that are out in the industry just didn't feel sufficient".
Evals are tiered: must-pass, tradeable secondary, then feature level. Evals are grouped into safety (prompt injection, general harm), capability (browser use, task completion), behavioral (follow-up questions, intent understanding) and use case. Priority evals like safety must always pass, secondary ones can be traded off slightly, and teams add feature-level evals.
PMs, engineers and researchers all write evals, with agents helping. Muse has too many to count.
Humans handwrite the system prompt and behavioral spec. A recent Muse rule says no part of the system prompt may be AI-written. People also review eval data to confirm its prompt distribution matches real product use.
Mona says issues appear whenever the team skips that review.
Agent swarms generate multi-turn data and find real production issues. Meta AI conversations were mostly single turn, so there was no data for long-running multi-turn evals. Agents swarm over conversations and synthetic data and return formatted dumps of what tasks people did and what worked or failed; the team swarms every new model and any suspected issue.
One suspected problem was swarmed, confirmed as real and given an owner. Findings "feed into our evals".
PMs read specific production examples, not just dashboards. Swarm output shows concrete cases, such as 10 people asking to shop on a site and getting blocked, which lets the team target fixes precisely. Mona makes all her PMs do this to build empathy.
"if you don't have that empathy and you can't feel the pain, you will never develop good products."
Agents in a group chat taught the team about discretion. Early on, two employees' agents shared a group chat and one told the other person its owner had been to the doctor. The agents later had a passive-aggressive fight, and the team decided multi-agent chat was not ready to ship.
"we learned about discretion the hard way"
A throwaway mascot beat the corporate logo because Muse belongs to the user. Lead designer Alex Cornell made Jolly with a one-shot prompt, named after his nickname for his daughter. A replacement logo was ready, but about a week before launch Zuckerberg's love of the avatar, team resistance and employee affection kept Jolly, and users can now make their own avatars.
Mona calls Jolly "our cockroach" because every few months someone said they couldn't launch with a stuffed animal at the top. Added animations like headphones and a laptop built more system around it. "When you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo."
An early principle settled design fights instead of opinion. The team stated early that the agent works for and is owned by the user, with guardrails and a charter from Meta. Once that existed, a corporate logo clearly violated it.
Mona tells PMs they have agency to push back on defaults: "it's not a random rationale".
Chat bubbles make the agent feel like messaging a person. While most AI products moved to large canvas, essay-style answers, Muse chose chat bubbles. That constrains long answers and forces choices like how to break up thoughts and whether a carousel feels odd.
The choice was controversial inside Meta and the team is still working through the downstream effects.
One main chat shipped first; side chats came only after demand. The team argued an assistant with long-running context should be one conversation, like a personal assistant who remembers everything. Employee reaction was the most polarizing of the first release, and side chats were added once demand for mental separation was clear.
Mona still uses only the main chat.
Word of mouth activates users that advertising could not. Mona watched a power-user mom pass her phone around a group of parents and scroll through how she used Muse; two of them downloaded it on the spot. The team tries to make it easy for users to share real stories and examples.
Activation is now the hardest problem. "the most effective lever out there is word of mouth."
Starter packs in an ideas tab fix first-prompt freeze. Most first prompts are "hi" or "what's the weather". The ideas tab offers packs, such as creating a notification cron or building a dashboard, where the user just says "let's do it" and Muse starts work.
Mona says it has gone relatively well, but people still struggle to imagine what the agent can do.
Later adopters get different onboarding than early adopters. Early adopters knew agents and wanted Muse to be safer and push boundaries; later cohorts ask how it differs from AI they already use. Muse now serves onboarding variants based on where users come from and how exposed they are to agents.
The team says it has not figured onboarding out yet.
Shopping became a top use case nobody predicted. Mona expected productivity, health and personal finance, others expected home automation, but dogfooders started shopping. Because evals existed per use case, the team refocused and added shopping evals.
Home automation, big in the OpenClaw days, did not take off with the broader base. One engineer shopped so much Mona joked about taking her credit card away.
Under 100 people and four PMs built the whole product. The first team was about 10 people, mostly engineers plus designer Alex Cornell, and the effort stayed under 100 including all functions. A PM joined an area once it was headed for the product or launch.
"communication management and decision making had to be very very tight"
Engineers prototype from a kernel of an idea before specs. Many features began with one or two engineers taking an idea away for a few days and returning with a prototype. Cornell built the first Muse prototype, a chat thread and a feed, in days; it was shown to the board after being fixed minutes before the meeting.
Mona says this needs a talent-dense team with product sense and a lot of trust: "You got to let the engineers cook."
Reviewing working prototypes replaces slide reviews and avoids sunk cost. Leaders reviewed phone prototypes in small rooms, and the team could say "I hate this" without months of engineering behind it. Mona can now get to an idea in a 24-hour cycle, sometimes building it herself.
She also uses AI to build internal tools that never got prioritized.
Product taste is the top quality Mona hires PMs for. When building is easy, shipping AI slop is easy too, so PMs earn their place as the voice of the customer holding a high bar. Execution still matters, but taste comes first.
"product taste now is the most important thing."
AI assistantsAn AI research and product company behind ChatGPT, Codex and the Dots agent.
AI assistantsOpenAI's AI assistant app used for chat, plugins and agent work.
AI assistantsAn AI company that builds the Claude models and products such as Claude chat and Cowork.
AI assistantsAnthropic's AI assistant and model family, used for chat, coding and agentic work.
AI assistantsGoogle's family of frontier AI models and the assistant app built on them.
AI assistantsA voice AI company with text-to-speech, dubbing, speech-to-text and music models, plus consumer and creator apps.
AI assistantsGoogle's AI notebook for working with your own documents and sources.
AI assistantsA voice-first AI companion app with an animated alien character that remembers your conversations.