Meta Muse: the AI agent people actually love
Alex Cornell explains how Meta's Muse was designed as a messaging-style entity rather than a software tool, and how he uses it to turn neglected personal data into useful, delightful artifacts.
Watch on YouTube ↗Charlie's summary
- ▶ 3:34
Design the agent as an entity you exchange messages with, not as a traditional AI app
Alex says the framing that clarified almost every decision was imagining the interface as an exchange between you and an entity. That is why the character sits at the top of the screen and why messages have bubbles, which help when an agent messages you proactively at different times. Because it can work in the background, showing a status of when it is working became “one of the most important decisions.” Ask how you would handle it with a real person.
- ▶ 7:06
Use progressive disclosure so normal people see simple status and engineers can drill down to the code
Alex layers status. The top says the agent is working, then an emoji says what it is working on, then a tap shows everything it has done, and a further tap reveals the curl and code for auditing. Chain of thought can get pretty loud in the core chat, so he asked whether it needs to be there for everybody. Probably not. Clarvo adds that narration often patches over latency, and Muse is simply fast.
- ▶ 9:40
Acknowledge hard requests quickly like a person would, then rely on status for long tasks
Alex says that when you ask a person for something hard, they don't ghost you until it's done. They say they'll do it and come back. Muse similarly acknowledges a complex request with a reaction or an initial message, then goes off to work. For long tasks the status display shows what is going on, so you can check whether it is stuck or going down the wrong path and never feel left alone.
- ▶ 11:11
Generative UI works best on fixed rails: stable primitives with generated content inside them
Clarvo praises Muse for keeping core primitives stable (avatar, status, ideas, feed) while generating what fills them: a fresh set of ideas, the right emoji for each status, and a video of the agent typing. Alex says it is more on the rails than fully generative. If you turn your avatar into a tomato, you see a tomato typing. Ideas based on what you've actually discussed, like golf, are far more useful than generic suggestions.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.