Agent UX design: For teams building AI agent products
An agent product asks people to hand over work. Whether they do depends on the model, and just as much on what the interface lets them do: predict what the agent will do, check what it did, and change it afterward. From 2022 to 2025 I was principal designer on Deva, a platform for creating and training AI agents, and 10Planet, the protocol underneath it. This is that work, for teams building their own.
To start, email andriy.dsgn@gmail.com with a few lines about the agent and what people are asked to hand over.
Signs your agent product needs it
- Users try the agent once and go back to doing the work themselves.
- People cannot tell what the agent will do before it acts.
- The agent ships at full autonomy because the demo looks better.
- Memory is invisible: nobody can see what the agent knows, or change it.
- Support tickets ask why the agent did something.
What I design
- The objects before the screens. What an agent is, the sources it learns from, who may talk to it, and the record of who contributed what. On Deva the screens were rebuilt several times; those objects changed much less.
- Teaching. Ways for someone who is not an engineer to give the agent knowledge, and a private stage to train it before it goes public.
- Memory you can see and scope. Every piece of memory with a place in the interface, a visible source and a way to change it. On Deva, three switches on every note: Use for Responses, Show in Retrievals and Publish.
- Autonomy per action. Trust extended in steps, set for each action rather than for the whole agent, with approvals where a mistake is costly.
- Execution, governance and interface kept apart, so each can change without dragging the others along.
- Measuring delegation: tasks handed over, finished and corrected.
How it runs
- Scoping. A one-week Product Diagnosis of the agent and its users, credited to the project.
- Object model. What the product is made of, before any screen.
- Flows. Teaching, delegation, approval, correction, with every state.
- System. A design system tied to the domain logic, so new features start from rules the system already knows.
Or as fractional product design, when the agent needs a design owner for months.
What the work produced
| Product | What it was | Measured or delivered |
|---|---|---|
| Deva | A platform for creating, training and working with AI agents | From the first concept to a beta in 15 months. In the beta, tasks took 30% less time in usability tests; daily active users grew 25% and retention 20% |
| 10Planet | The protocol under Deva | Agents, memory and the design system mapped to protocol logic |
| Fully AI | An AI agent product for businesses | A UX audit, then a design system and a dashboard redesign, in one month |
Measured by the teams, not by me, and reported to me by the people who ran the numbers. Each compares with the version before the change: Deva’s in its beta, Equi’s while I was on the team, 2021-2022. My part in each is on the case page.
Questions teams ask
What kinds of agent products do you design for?
Products where people hand work to an agent and have to trust it: agents that act or answer in someone’s name, assistants inside a product, platforms where people build their own agents.
How much does it cost?
A project is quoted after a one-week scoping Diagnosis, from $900 and credited to the project. As fractional work, from $8,000 a month.
Do you need to understand our model and stack?
I need to understand what the agent can and cannot do, and what it costs when it is wrong. I read the documentation and talk to the engineers before I design.
Can you design only the approval and control flows?
Yes. Approvals, undo and autonomy settings are a common first project.
Do you work with teams in other countries?
Yes. I work remotely, with teams in different countries and time zones.
Other ways I can help
- Fractional product design, from $8,000 a month: a senior designer, part-time.
- Design systems, from $700: a Consistency Review, then a system agents follow.
- Product Diagnosis, from $900: one week, a written plan: keep, cut, change, next.
Further reading: Designing agent products. All services, with the situation each one fits: Services.
Building an agent people have to trust?
Send a few lines about the agent and what it does for people. If it’s not a fit, I’ll say so.