Distyl is engineering-heavy hard tech. Brilliant multi-agent systems doing real enterprise work. No design function, no visual language, no component system, no brand. I was the first designer.

The real problem wasn't the missing design system. It was that the product worked and nobody outside the building could read it. Enterprise stakeholders had to act on agent outputs they couldn't trace. Distributed logic, non-deterministic behavior, no way to see what the system was doing or why. The agents were trustworthy. The experience wasn't.

First instinct was the obvious one: build the system, ship the components, give it a coherent surface. I did that. It wasn't enough. A clean UI on top of an opaque agent doesn't make the agent legible. It just makes the confusion better-looking.

So I reframed. The job wasn't to design screens. It was to design the operator's mental model of what the agents were doing. I named the cost of getting it wrong: the Context Tax, the re-orientation users pay every time a system fails to hold what it should. That framework became the filter for every decision after.

The decision worth narrating: I made agent state visible at every step instead of hiding it behind a clean result. Conventional UX says reduce, simplify, hide the machinery. For non-deterministic systems that's exactly wrong. Operators trust what they can trace. So the work exposed the reasoning path rather than concealing it. Plus the foundation underneath it: visual language, components, brand, all built zero to one, all in service of that one principle.

What it confirmed: in AI products the bottleneck is almost never the model. It's whether a person can bring themselves to the work. Legibility is the design problem, and most teams find that out too late.