Key Takeaways

  • Alex Cornell, design director on Meta's Muse, argues that building consumer agents requires unlearning deterministic UI design where engineers and designers control every pixel.
  • Because each user's setup and outputs in Muse differ completely, design work shifts from policing individual screens to creating underlying scaffolding and system architecture.
  • Consumer-facing AI agents work best when styled after messaging platforms instead of developer tools, using progressive disclosure status systems to prevent interface clutter.
  • Cornell advises designers to accept variable visual outputs and resist the urge to hand-craft every artifact rendered by non-deterministic models.

The Death of the Deterministic Canvas

For decades, software design operated on a simple contract. A designer mocked up an interface in Figma or Sketch, passed it to an engineer, and expected every customer to see the exact same layout. As Cornell told Claire Vo: “for through my whole career you would come to a feature and you would be able to think about it and then design it and kind of know that like on any given customer screen it's basically going to work like this and look like this because that is the way that this thing has been coded.”

Non-deterministic systems break that contract completely. When an AI agent generates the content, structure, and next step in real time, no two screens match. “My Muse setup is completely different than yours is completely different to somebody else's,” Cornell noted. “They're so personal and they can do so many different things that you really kind of have to think more about the underlying foundation and the system, the architecture, the scaffolding that's in place, and less about like well what about this one little thing here, is it going to be just so.”

Trying to enforce pixel-perfect control over an agent's visual output creates friction. The software evolves faster than static design systems can keep up. If you spend forty hours tuning the margins on an artifact card that only three users will see in that exact configuration, you waste time.

Designing Scaffolding Instead of Outputs

The solution is building structures that absorb unpredictability. On Meta's Muse, that meant taking cues from messaging apps rather than developer dashboards. Developer tools expose dials, raw JSON trees, and dense configuration panels. Consumer users do not want that complexity. They want a conversation with clear boundaries.

To keep conversational feeds clean, Cornell implemented a progressive disclosure status system. Instead of overwhelming the screen with step-by-step model reasoning, the interface tucks execution details behind concise indicators until the user asks for more context. The system communicates state without burying the feed in technical logs.

This structural approach forces a mental reset for visual perfectionists. Generated artifacts will sometimes look imperfect or raw. Cornell admitted that letting go takes discipline: “Actually one of the things that's hardest for me sometimes is looking at those artifacts that we were looking at. I kind of have to turn off my brain sometimes design-wise because you can, if you want, go and spend all the time in the world to make it look exactly as you want, but I've come to kind of view it a little bit more freely and it's kind of liberating in some ways.”

What to Do With This

Open your product and identify the screen where your AI model outputs dynamic content. Strip away any rigid layout rules or multi-step form fields that force the output into a predetermined container. Replace them with a messaging-style container and a progressive disclosure status tracker that summarizes agent work in one line until clicked.