Most agentic UI is designed by technical early-adopters, for technical early-adopters. The people building the future of AI interfaces tend to be the same people most comfortable with a blank prompt and a blinking cursor. So that is who the agentic UX quietly gets tuned for — someone who already knows what to type.
At Tessa, building ROCCO, we didn’t have that luxury. And I’ve come to think it was the best constraint we could have had.
Our users run municipal waste collection. Many of them spent decades driving the truck before they ever managed the routes. Their expertise is entirely real and entirely physical — hills, alleys, which streets flood, how long a heavy Monday actually takes. A blank chat box is not a friendly thing to hand that person. It’s a test they didn’t sign up for. So we stopped treating “make it powerful” and “make it simple” as opposite ends of a slider, and started designing the agentic UI for the least technical person who would ever depend on it.
The contrarian bet: design down, not up
The instinct, when you’ve built a genuinely clever agent, is to let it show its work — explain every option, surface every setting, prove how smart it is. That’s the default in most AI product design, and it’s backwards.
The contrarian move is to design down. Assume the person in front of the screen has never used software like this, doesn’t want to learn your vocabulary, and is one confusing panel away from going back to the whiteboard. Then build for that person. Not a dumbed-down version for them and a “real” version for everyone else — one interface, aimed at the floor.
That reframing matters because “accessible AI design” usually gets filed under compliance. Here it was the product strategy. And it turns out designing agentic UI for non-technical users is also the fastest path to enterprise AI adoption, because the person who has to actually live in the tool every day is rarely the person who bought it.
The four agentic UI rules that fell out of it
Designing for our users produced a short, stubborn set of rules. None of them are clever. That’s the point.
- Plain English, always. On any control we say three things and stop: define the term, name the safe default, and quit. Edge cases and “what happens if I tick this” go to the agent, behind a one-line “ask ROCCO” pointer — never crammed into the panel.
- Never show a database ID. People see route names and street names. The internal numbers stay internal, forever. Nobody should have to learn our schema to do their job.
- Voice input, because their hands are busy. Tap the mic to toggle it on and off — no press-and-hold gymnastics while you’re standing at a truck.
- Short messages. The app says the one thing you need to know. The detail lives in the conversation, if and when you ask for it.
Look at what those rules have in common: every one of them removes something the agent could have shown off. That’s the whole discipline of good agentic UX — the agent’s job is to keep the surface calm and say more only when asked.
Human-in-the-loop, without the homework
None of this makes the agent less capable. ROCCO still does the hard work — sequencing routes, weighing tradeoffs, flagging what looks wrong. The human stays in the loop for every real decision. We just refused to make staying in the loop feel like homework. Designing AI agents this way means the person keeps authority without needing a manual to exercise it.
Here’s the part I didn’t expect. Design your agentic UI for the least technical person who will ever touch it, and it gets clearer for everyone — the power user, the buyer, me. Nobody has ever complained that the interface was too obvious. Plain language, calm surfaces, and no jargon are not accommodations you make for the few. They’re just better AI UX.
“Simple enough for someone who’s never done this” isn’t a downgrade. It’s the spec.
