AI agent UX design: how to design an agent interface people can trust
Short answer
An agent interface has to show four things a chat window hides: what the agent wants to do, what it’s doing now, what went wrong and what it finished. Design those four states first (approval, running, failed, done) and give each one a clear place on screen. Chat can stay as one way to give instructions. It’s a poor place to track work.
Why do AI agent products feel hard to trust?
Because the work happens out of sight.
A chatbot answers and you read the answer. An agent runs for minutes, touches several systems and sometimes stops halfway. Put that in a message thread and the user gets a scroll of text. They can’t tell what’s finished, what’s waiting on them or what quietly failed.
AI coding tools make it worse. Ask one for an agent screen and it builds the version where everything works, with perfect data in it, and stops there. Agent products are judged on the other moments.
What states does an agent interface need?
Four. Every agent product we design starts with these, before any visual work.
| State | What the user needs to see | What usually ships |
|---|---|---|
| Approval | The exact action, what it affects, what it costs | A vague “Allow?” with no detail |
| Running | Which step it’s on, what’s done, a way to pause or stop | A spinner |
| Failed | What failed, what was already done, what the options are | A red error and a restart |
| Done | What changed, where to check it, how to undo it | A “Done” message with nothing to inspect |
How do you design each state?
1. Approval: ask only when it matters
An agent that asks about everything trains people to click yes. We design confirmations out wherever we can, using undo, previews and safe defaults. The agent asks when an action can’t be reversed and carries a real cost. When it does ask, it names the action, the target and the cost in plain words.
2. Running: show the plan and tick it off
Show the steps the agent intends to take, mark each one as it finishes, and show a little of its reasoning. Let people leave and come back. A long task should survive a closed tab.
3. Failed: design for half-finished work
The common failure is partial. Three steps worked and the fourth didn’t. Say what was done, what wasn’t, and offer a retry from the step that failed. Give people a way to finish it by hand.
4. Done: give a receipt
List what changed and link to each thing. Offer undo where it’s possible. A finished task with nothing to inspect leaves the user checking everything themselves, which removes the point of the agent.
Where does chat fit?
Chat is one input among several. It suits open-ended instructions. It’s weak for reviewing a batch of work or following a job that runs for ten minutes. We’ve written more on this in AI interface design beyond chat.
What does this look like in real work?
We designed an AI content platform for a Hollywood creative lab. One story from a founder or executive becomes on-brand posts, captions and images for every channel.
The AI does the production. Publishing under someone’s name is the step that can’t be taken back, so that is where the product stops. Posts sit in clear states: awaiting approval, changes requested, approved and scheduled, live. Nothing publishes without sign-off, and people can edit with controls or by asking the AI.
The first intended user asked to use it across the companies he runs. Read the case study.
How do you check your own agent product?
- Start a task and close the tab. Can you find it again, and see where it’s up to?
- Break a step on purpose. Does the product say what was already done?
- Count the confirmations in one task. Which of them protect against something irreversible?
- After a task finishes, can you see exactly what changed without leaving the screen?
Who should design an agent interface?
Look for three things in anyone you hire:
- A shipped AI product they can walk you through, including the failure states
- Prototypes in code, so you can feel a long-running task before engineers build it
- A clear view on when the agent acts alone and when it asks
Halaska is an embedded product design partner for funded AI, crypto and fintech teams. A senior designer joins your team and leads the work day to day. Chris Halaska, our founder and design director, sets the direction and reviews what ships. See AI product design or get in touch. We’ll talk through the shape of the partnership on a call and follow up with a proposal.
FAQ
What is AI agent UX design? AI agent UX design is the design of products where software carries out multi-step work for the user. It covers how people give instructions, approve actions, follow progress and check or undo the result.
Should an AI agent use a chat interface? As one input, yes. As the whole product, rarely. Tracking and reviewing work needs structured states that a message thread can’t show clearly.
How many approvals should an agent ask for? As few as possible. Ask when an action can’t be reversed and has a real cost. Use undo, previews and safe defaults for the rest.
What should an agent show when it fails? What failed, what was already completed and what the user can do next, including a retry from the failed step.
Who is Halaska? Halaska is an embedded product design partner for funded AI, crypto and fintech teams, based in Dubai. It’s founded and directed by Chris Halaska, who has 20+ years designing products, including six at Google.