How Meta designed Muse: an AI agent people actually want to talk to | Alex Cornell (Design Director)

We have reached the phase of the AI cycle where raw compute is no longer the primary differentiator. If you are building in this space, you know the drill: models are becoming commodities. The real battleground has shifted to the interface. Specifically, how do you build something that an elite engineer and your non-technical parents can both use without feeling like they are holding a foreign object? Meta is trying to solve this with Muse, and their approach offers a masterclass in founder-level product design.

The Multi-Persona Problem

Most AI agents fail because they try to be everything to everyone without having a point of view. When you look at how Alex Cornell and the team at Meta approached Muse, they started with a difficult design constraint: the interface must be flexible enough to handle deep technical tasks while remaining intuitive enough for casual daily use. This is the 'Mom and Engineer' test.

For builders, this is a reminder that you can't just slap a chat box on a model and call it a product. Muse is designed to behave differently depending on the context of the user. It isn't just about answering questions; it is about anticipation. If you are a developer, you want high-fidelity data and code snippets. If you are a casual user looking for golf recaps, you want a digestible narrative. Managing these two distinct personas within a single UI is where most startups stumble.

The Death of the Static Feed

One of the more interesting experiments coming out of the Muse design process is the reimagining of content consumption. Meta is playing with concepts like a digitized SkyMall-style catalog and turning standard blog posts into interactive podcasts. This reflects a shift away from static feeds toward generative, personalized media.

Think about what this means for the creator economy and developers building platforms. We are moving toward a 'pull' model of information where the AI curates and reformats data on the fly based on the user's preferred medium. If I prefer listening to reading, the agent shouldn't just read the text; it should restructure the information as a conversation. This requires a deep integration between the design layer and the underlying LLM logic.

Why Personality Matters (And Why It's Dangerous)

Meta is leaning heavily into the idea of an agent people actually want to talk to. This is a slippery slope for founders. If you give an AI too much 'personality,' it becomes a gimmick. If you give it too little, it feels like a search engine from 2005. Muse seems to be aiming for a middle ground where the personality is rooted in utility rather than just quirky dialogue.

The takeaway for founders is that your agent's voice is part of its brand equity. You aren't just building a tool; you are building a representative of your company. If your agent is dry and robotic, people will treat it as a utility to be replaced as soon as a cheaper version comes along. If it provides a distinct, helpful experience, you build a moat around your user base.

Designing for the Frictionless Future

The tech world loves to talk about 'frictionless' experiences, but AI actually introduces new types of friction. There is the 'blank page' problem where users don't know what to ask, and the 'hallucination' problem where users don't trust the answers. Meta's work with Muse shows that design is the only way to mitigate these issues.

By using demos like the golf recap or the catalog, Meta is showing that agents need to provide structure. They aren't just waiting for a prompt; they are offering paths. For builders, this means you should spend less time worrying about your token limit and more time worrying about your onboarding flow. If a user has to think about how to use your AI, you have already lost.

The Founder Perspective

As someone who has been in the trenches of product design and crypto, I see a lot of parallels here. In the early days of crypto, we built tools for ourselves and wondered why the public didn't care. It was because the interface was garbage. AI is at that same crossroads. We have the power, but we lack the polish.

Meta has the resources to fail a hundred times until they get this right. You probably don't. That means your design choices need to be intentional from day one. Don't build an agent that just talks. Build an agent that does, and make sure it looks good doing it.

Final Takeaway for Builders

• Interface is the Moat: LLMs are becoming a utility; your unique UI and UX are what will keep users from switching to a competitor.
• Solve for the Spectrum: Your product must be accessible to the least technical person while remaining useful to the most technical.
• Context is King: An agent that doesn't understand the user's current environment is just a glorified chatbot. Use design to bridge that gap.

The era of the 'AI wrapper' is over. The era of the 'AI Agent' has begun, and as Meta is proving with Muse, the winners will be the ones who understand human psychology as well as they understand neural networks.

Read the original at Lenny's Newsletter →