Inside Meta's Muse: Building an AI Agent

On a Saturday in early February, Meta product manager Mona Sarantakos's children left for a ski race, her husband went with them, and she sat down with a Raspberry Pi to talk to an open source AI agent called OpenClaw. By Sunday she had given it access to her email, her calendar, and her home automation. Her husband came home and asked what she was doing. "That is crazy," he said. "That is so not safe." She kept going anyway (04:10).

The payoff came Monday morning. Sarantakos says she is bad at tracking school newsletters and Spirit Week schedules. That morning her agent had already pulled the week's highlights into one note, built from a prompt of only a few words (05:06). "It felt pretty magical," she says, "with such little direction from me" (05:32). That weekend became the seed of Muse, the personal AI agent that Meta's Superintelligence Labs, the company's AI research division, shipped seven months later (09:48), and which Meta CEO Mark Zuckerberg has hinted is the fastest growing AI consumer product of all time (01:32).

Sarantakos, now VP of product for Muse, tells the story to Lenny's Podcast as a case study in restraint. The team had working prototypes well before launch. What took time was trust. "Trust is a must," she says of the internal mantra that governed the build.

"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 gonna use it." (Mona Sarantakos, [12:33])

So the team spent most of its seven months on reliability, safety, and security rather than new features, widening its testing every time a capability, like giving the agent a browser, opened a new security risk (11:49). It leaned on alpha testers and internal dogfooders through the summer, and ran what it calls agent swarms, fleets of AI agents that scan production conversations, to generate the kind of multi-turn usage data that did not exist anywhere else because nobody had built a long-running personal agent before (39:55).

That same caution almost killed the thing that makes Muse recognizable: a small, soft avatar named Jolly. Lead designer Alex Cornell built it as a placeholder in the first prototype and named it after his daughter's nickname, Jollybug (24:14). For months, people inside Meta assumed marketing would swap it for something more corporate. "Every few months somebody would be like, 'There's no way you're gonna launch this product with a stuffed animal at the top,'" Sarantakos says (00:52). The team kept arguing it should stay, and about a week before launch, Zuckerberg sided with them (26:58).

"When you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo. That breaks the spell completely." (Mona Sarantakos, [26:33])

The decision reflects a principle Zuckerberg pushed early: Muse exists to help people reach their goals, not just complete tasks (21:04). Once the team internalized that, they built a "Goals" tab and started designing for what they call dreaming and studying, the idea that a 24 hour agent can think about a user's priorities even when nobody is actively chatting with it (21:43). It also explains why the product uses ordinary chat bubbles instead of a document-style canvas, so talking to Muse feels like texting a person rather than opening a dashboard (52:05).

Some predictions turned out wrong. Sarantakos expected personal productivity, health, and personal finance to be the breakout use cases, with home automation close behind. Instead shopping became one of the largest categories, a genuine surprise that forced the team to build new evaluations specifically for it (32:12). One engineer on the team, Christine Awad, became the resident stress test. "I need to take your credit card away," Sarantakos jokes, describing mornings spent reading Awad's shopping bug reports (33:52).

Handwriting in an Automated World

The most counterintuitive rule to come out of the build is that a company built on AI insists a human write the AI's own instructions. Muse runs on more than 100 individual evaluations, or evals, organized into tiers: priority evals for safety and prompt injection, capability evals for task completion, behavioral evals for personality, and use case evals for scenarios like shopping (43:05).

"The system prompt has to be handwritten. You cannot use an AI to write a system prompt, any part of it." (Mona Sarantakos, [45:10])

The same rule covers the behavioral spec and the prompt distributions used to test the product. The logic is simple: if the instructions that shape an agent's judgment are themselves machine generated and unreviewed, nobody can be sure they still match what the product is supposed to do.

Sarantakos has been a product manager for 20 years, and she argues the job itself has shifted under her. Engineers can now prototype an idea in days instead of months, which means arguments that once happened over slide decks now happen over a working build on someone's phone, Zuckerberg included (17:45). That speed changes what a PM is actually for.

"Product taste now is the most important thing. Having, like, that kind of, like, high taste, high bar, like, product instinct." (Mona Sarantakos, [65:37])

Execution still matters, she says, but engineers with strong product sense can execute well on their own. What a PM adds now is judgment: the willingness to say a feature an engineer loves will not land with ordinary users, and the empathy to know why. The agent swarms exist partly for that reason. They do not just return metrics; they return stories, and stories build empathy for a user's actual frustration (41:34).

Small Teams, Bigger Questions

Muse was built by a startup sized crew inside one of the largest companies in the world. The first version of the team was about 10 people (75:12); the core group never grew past 100, with only four product managers covering the entire build (76:04). Keeping it small, Sarantakos says, is what let decisions stay close to the people who understood the product best, though it also meant little sleep and no slack when something broke.

What comes next worries her more than it excites her, in a useful way. Muse now manages real responsibilities for real people: a single mother's medical appointments, a parent's kid-sports schedule, a household's shopping (79:13). Sarantakos wants the next version of Muse to need less supervision, moving past what she calls "babysitting" an agent toward one that can intuit what a user wants and act on its own (82:19). She also expects the product to leave the phone screen entirely, migrating into glasses and other ambient hardware, "in your glasses, on your charm devices" (81:22).

It is a strange arc for someone who grew up changing television channels by walking up to the set and turning a dial. Sarantakos is 44, and she says plainly that she doubts she will work on anything this personal again, not because what comes after is smaller, but because this product was built, in the beginning, purely for the small team that made it, before anyone else was watching. That it now sits in millions of pockets is, by her own account, still closer to vertigo than victory, and she seems to prefer it that way.