Marcus Andrey Study
Ler em português Library

AI for UX · Step 5 · 6 min read

Agentic design

How to design AI features and agents people can trust: the spectrum of autonomy, principles for AI that acts on people's behalf, the UI patterns that make agents visible and controllable, and the notice each pattern depends on.

An AI feature suggests; an agent acts. When software starts doing things for people (booking, filing, changing data, sending messages), the design question changes from "is the answer good?" to "do people know what it's doing, can they stop it, and can they undo it?" This note is about designing for that: how much autonomy to give, which principles keep people in control, and which interface patterns make an agent visible.

Six railway signals, one for each kind of notice an AI agent sends to the screen: started, step, data changed, using a tool, needs approval, error. Each one is a state a designer has to design. Point at a signal to raise it. Illustration: Marcus Andrey, drawn with Hairline (MIT, Lucas Marques).

What you'll be able to do after reading:

  • Choose an autonomy level for an AI feature based on risk, not on what's technically possible
  • Apply eight principles for AI that acts on people's behalf
  • Use the core UI patterns for agents, and name the notice each one needs from the agent
  • Plan research that tests whether people trust an agent the right amount

1. Two meanings of "agentic design"

  1. Designing for agents: the user experience of AI features and agents inside a product, for example an assistant that reschedules deliveries when a truck breaks down.
  2. Designing with agents: designers directing AI agents that do multi-step design work. See Generating design with Claude Code and Building an AI agent roadmap for your UX team.

This note is mostly about the first. NN/g's working definition is useful: an AI agent is a system that pursues a goal by planning and taking actions with tools, with some independence (NN/g).

The cockpit of a C-141C transport aircraft, with screens, autopilot controls and two control yokes.
Aviation has designed for levels of automation for decades: the autopilot flies, the crew sets the goals, watches the instruments and can take over at any moment. Photo: TSgt Lance Cheung, US Air Force, Wikimedia Commons, public domain.

2. A spectrum of autonomy

1Suggest

AI recommends; the person decides and acts.

"These 3 deliveries may be late"
2Draft

AI prepares; the person reviews, edits, confirms.

A pre-filled form from a scanned invoice
3Act with approval

AI plans and acts after the person approves.

"Reschedule 12 deliveries. Approve?"
4Act and report

AI acts within limits and reports; the person can undo.

Reschedules when the cost difference is small, with a daily summary
5Autonomous

AI acts and corrects itself; people set goals and rules.

Rarely right for high-stakes work
Five levels of autonomy. Choose by risk and earned trust, not by what is technically possible. Diagram: Marcus Andrey, CC BY 4.0.

Design rule: choose the level by risk (cost of an error, whether it can be undone, legal exposure) and by earned trust, not by technical possibility. Move up one level only with evidence that it's safe.


3. Eight principles

  1. Make the AI's role clear. What it can and can't do, and where its information comes from.
  2. Show confidence and reasons. "Taken from the invoice, page 2; not sure about the tax code." See Designing for uncertainty.
  3. Keep people in control. Preview, confirm, undo. Clear points where a person takes over. See Human-in-the-loop.
  4. Design for being wrong. Easy correction, and corrections that feed back. Errors are a when, not an if.
  5. Show progress and state for long tasks: what's done, what's next, what needs me.
  6. Keep an audit trail: who or what did what, when, based on what.
  7. Fit the work people already do. A chatbot is rarely the best answer; a better default, a bulk action or an inline suggestion often is. See Conversational and generative UI.
  8. Hand over gracefully. When the agent is stuck or unsure, it passes the task to a person with context, not a dead end.

These echo the 18 Guidelines for Human-AI Interaction from Microsoft Research (Amershi et al., CHI 2019), still the best-known checklist in the field.


4. UI patterns, and the notice each one depends on

An agent can only be shown well if it tells the screen what it's doing. Each pattern below depends on a specific kind of notice (an "event") from the agent. If the engineering team doesn't send that notice, the pattern can't be built. The next lesson, Conversational and generative UI, explains these notices in detail.

PatternProgress indicator

What is done, what is next, what needs me.

Needs: run started · step started/finished
PatternPlan preview

"I will do these 5 steps."

Needs: the plan, before any action
PatternInline suggestion and diff

A changed value, marked, with accept and reject.

Needs: state changed
PatternSource citations

Where an answer came from.

Needs: tool call and its result
PatternApproval dialog

A clear decision point before a risky action.

Needs: needs approval (an interrupt)
PatternActivity log

What the agent did, when and why.

Needs: activity and the full run history
PatternUndo

Put things back as they were.

Needs: a snapshot of the state before
PatternGuardrail and error message

"I can't do that because…", with alternatives.

Needs: error, with a reason
Agent UI patterns and the notice each one depends on. If the agent does not send the notice, the pattern cannot be built. Diagram: Marcus Andrey, CC BY 4.0.

5. Research and evaluation

  • Wizard-of-Oz tests before building: a person plays the agent behind the scenes to test the interaction.
  • Test with realistic failures: include deliberately wrong suggestions and see whether people catch them. NN/g found that AI chatbots tend to discourage error-checking (NN/g). See Automation bias.
  • Measure acceptance, correction, time saved, errors introduced and trust over time. See Measuring AI features and Evals for designers.

6. Governance angle

  • In the EU, AI that allocates tasks based on individual behaviour or personal traits, or monitors and evaluates workers, is a high-risk use under the AI Act, with obligations from 2 December 2027 as amended (Annex III). Flag such agents early. See AI governance and data protection.
  • Transparency: people must be told when they're interacting with AI, unless it's obvious (Article 50, since 2 August 2026).

7. Vocabulary

Term Meaning
AI agent A system that pursues a goal by planning and taking actions with tools
Autonomy level How much the AI does without asking
Human-in-the-loop A person reviews or approves at key points
Plan preview Showing the steps an agent intends to take, before it acts
Audit trail A record of what was done, by whom or what, and why

8. Try it (no code needed)

  1. Pick an agent idea in your product ("reschedule my deliveries when a truck breaks down").
  2. Place it on the autonomy spectrum (section 2) and write why.
  3. Sketch three states: the plan preview, the approval moment, and the activity log after it's done.
  4. For each state, write the notice the agent must send for it to exist (section 4).

Reflection questions: Which of your product's actions would you never let an agent take without approval? What would a user need to see to trust it one level more?


9. If you lead a team

  • Make sure every AI feature has a designer involved from the start.
  • Add AI patterns to the design system: suggestion chip, confidence badge, plan preview, activity log.
  • Create an AI feature review checklist from the principles above.
  • Agree a shared language with product and engineering: autonomy level, risk, reversibility, audit, notices.

Checklist

Learn more

Books, courses and videos to go deeper

Book links go to Amazon. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.

Conversational and generative UI · Designing for uncertainty · Evals for designers · Measuring AI features · Human-in-the-loop · Trust calibration · AI governance and data protection

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close