The world is too loud. Read what matters.

Lenny's Podcast

Meta's Muse: Waiting seven months because trust, once broken, cannot be recovered

After the team verified personal agents' potential with OpenClaw internally, they deliberately chose to wait seven months—because if something goes wrong early, users will never trust that agent again.

Consumer AIProduct managementMetaAI assistantAgent designPrototyping

The video won't play here. Listen to the audio instead:

First-hand account of how Meta transformed a ‘weekend toy’ into the year's fastest-growing consumer AI product in seven months—concrete details, plain speaking, and no padding.

The argument · tap a timestamp to hear it

4:06

A weekend's OpenClaw experience ignited the entire project

Nat Friedman messaged a few people on Thursday saying OpenClaw was changing his understanding of personal superintelligence. Over the weekend, Mona set up OpenClaw on a Raspberry Pi, connecting her email, calendar, and home automation. Monday morning, the system auto-generated a ‘weekly to-do’ summary that gave her an unprecedented sense of control with minimal prompting, compelling her to commit fully on the spot—even though the project didn't officially exist yet.

— Mona
12:11

Adding each browser feature meant discovering new security vulnerabilities to seal

The product initially lacked browser capabilities. Adding them created an entirely new category of security vulnerabilities. Rather than cutting features, the team fully sealed the attack surface opened by each critical capability—a direct cause of the seven-month delay.

— Mona
24:28

Mascot Jolly almost got replaced by the company logo

Designer Alex Cornell created Jolly Bot as a temporary character with a single prompt, meant to be replaced by the corporate logo before launch. The question kept being raised and rejected. The team's logic: talking to something this personal, you don't want to face a corporate logo—it would destroy the sense of connection. Jolly stayed, becoming the core symbol of the product's personality.

— Mona
31:33

Shopping use case was entirely accidental discovery, not predicted

The team initially prioritized personal productivity, health, and personal finance as core use cases, and expected home automation to scale—neither did. Shopping kept emerging during dogfooding, forcing the team to shift evaluation resources to support it. Shopping is now among Muse's most-used features.

— Mona
38:47

Sending out agent swarms to generate their own training data

The team lacked real conversation data from multi-turn, long-horizon tasks—different from the single-turn interactions that dominated earlier Meta AI and code systems. They deployed agent swarms to observe production conversations and generate synthetic data, producing regular formatted ‘dumps’ of ecosystem activity. This became the core data source for evaluation and enabled precise problem diagnosis.

— Mona
44:53

System prompts must be hand-written, cannot let AI ghost-write

Though the team extensively used agents for work, system prompts were entirely hand-written—AI generation was prohibited. Evaluation prompt distributions required manual review, and behavioral guidelines had to be hand-written as well, because they must precisely reflect the product's actual behavior. The team discovered that letting AI write them caused problems.

— Mona
51:59

Chat bubbles rather than canvas was a widely opposed bet

The entire industry—including Meta AI itself—was moving toward canvas-based immersive experiences. Muse went the opposite way, choosing chat bubbles, because personal agents should feel as natural as texting a person. This created a cascade of unforeseen problems: whether long responses should split across multiple bubbles, whether carousel cards belong in chat. The team admits these questions remain partly unsolved.

— Mona
1:16:26

Core team deliberately kept under one hundred people scale

The first version shipped with roughly ten people, mostly engineers. Even at launch, the core team was under one hundred. Only four PMs existed, with many modules having no dedicated PM—designers or engineers led decisions directly instead. Small size meant tight communication and fast decisions, but also almost no buffer: everyone chronically sleep-deprived.

— Mona

In their own words · checked verbatim

it was like such like such a small prompt that I had written. Right. Maybe like a few words. And the idea that I could kind of suddenly feel really productive and on top of my shit.

Mona5:06

There's a saying that we have internally, trust is a must. But really, like we were like in that, you know, it's like if we launch too early and we have an issue, like we have a bunch of issues, like that's it. Right. People will not trust this agent.

Mona12:11

Jolly is like our cockroach. It just like never dies. And every few months somebody would be like, there's no way you're going to launch this product with a stuffed animal at the top.

Mona23:28

when you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo. Right. Like that breaks the spell completely.

Mona25:29

I don't think I've ever experienced something so personal.

Glossary

zero basing
Starting from scratch rather than iterating on existing code or design.
agent swarm
Multiple agents deployed to autonomously observe production conversations and generate synthetic training data.
dreaming and study
Proactively reasoning about how to help users achieve goals without being directly prompted.

How to listen

Who it's for

Product managers and designers working on consumer AI products, and practitioners curious about how top-tier tech companies decide launch timing and personality-driven design.