Designing Meta Muse: Status, Personalization and Everyday Agents

Claire Vo interviews Alex Cornell, the Meta design director leading design on Muse, about a consumer AI agent and his own experiments with it. Vo is enthusiastic about the product, while Cornell works on it; their impressions are not an independent evaluation. The episode is sponsored.

Cornell says the team drew on messaging patterns because an agent can initiate messages and work in the background. A visible character and status indicator help people understand that work is underway.

The product surfaces a character, ideas and a feed instead of foregrounding software concepts such as files and diffs. Cornell describes a layered status: a simple working indicator, a short description of the activity, and more detailed traces for people who want to inspect them.

Vo says response speed affects whether users feel the agent is working or stuck. Cornell adds that a quick acknowledgement can be useful before longer work, with status details available when needed.

They distinguish a stable interface from generated content within it. Personalized ideas, avatar states and feed items vary, while the underlying surfaces remain designed components rather than an entirely generated interface.

Cornell demonstrates a shopping catalog built from saved Instagram material, with monitoring and price alerts. The example uses his personal saved data; the result depends on what access a user chooses to provide and what information is available.

For golf, he provides round and shot data and uses Muse to generate recaps, visualizations and practice feedback. He describes a long-term goal of a single-digit handicap, not a completed achievement.

Vo asks when to use the main chat versus side chats. Cornell describes the main chat as a continuous primary conversation and side chats as a way to scope or organize work; he says the distinction and organization still need improvement.

Other experiments repurpose old blog posts retrieved through the Wayback Machine into a morning audio reading and turn chess data into a game-by-game dashboard. These examples illustrate using an agent to reorganize existing personal material.

Cornell says designers must adjust to nondeterministic output: instead of controlling every pixel of every result, they need to design the underlying structure and allow useful variations. On prompting, he describes his own polite conversational style without claiming that it is universally optimal.