Oct 12, 2026 · 38 min
Meta designs Muse around conversation, not software
How Meta designed Muse: an AI agent people actually want to talk to | Alex Cornell (Design Director)
Muse shows how AI agents may need interfaces built for uncertainty, personal context, and ongoing collaboration rather than fixed workflows.
- 1Muse uses avatars, progress updates, and generative interfaces to make an agent’s behavior feel legible and approachable.
- 2Connected personal data lets Muse turn scattered saves, hobbies, and old writing into monitoring tools and interactive artifacts.
- 3Designing for nondeterministic agents means accounting for uneven results, latency, chat structure, and recovery when systems go off track.
Don't miss
Alex demonstrates how Muse transforms his old blog into a nostalgic, podcast-enhanced way to revisit writing from twenty years ago.
The brief
Alex Cornell, Meta’s design director for Muse, argues that consumer agents need human-centered primitives—avatars, status updates, and generative interfaces—instead of software concepts like files and threads.
Muse’s practical value emerges through personal context: it can organize Instagram saves into shopping ideas, monitor interests, compare prices, and help track long-term hobbies such as golf.
The standout demonstration turns Alex’s old blog into a nostalgic, podcast-enhanced way to revisit writing from twenty years ago, showing how agents can make personal archives useful again.
That flexibility creates a design problem: Muse’s outputs vary in quality, latency, and direction, so users need clear chat organization and ways to correct an agent without treating it like deterministic software.
The episode’s larger argument is that useful agents will be judged not only by what they can do, but by how gracefully they communicate progress, uncertainty, and mistakes.