Marcus Andrey Study
Ler em português Library

AI for UX · Step 5 · 7 min read

Conversational and generative UI

How to take AI beyond the chat box: why an agent must tell the screen what it's doing, how each of those notices becomes an interface state you design, how AG-UI (a 2026 open protocol) standardises them alongside MCP and A2A, and a no-code exercise to design the states yourself.

Most AI features still live in a chat window, beside the real work. The next step is AI that works inside the product: filling a form, updating a table, showing a plan, asking for approval in the right place. That only works if the agent tells the interface what it's doing, moment by moment. Each of those notices is an interface state, and designing those states is design work, exactly like designing empty, loading and error states.

What you'll be able to do after reading:

  • Tell apart conversational UI, generative UI and AI embedded in an existing interface
  • Explain why an agent's notices are interface states you have to design
  • Place AG-UI, MCP and A2A in one simple picture
  • Design the states of an agent-driven flow, without code

1. Three ways AI shows up on screen

Kind What it is Good for Watch out for
Conversational UI A chat: people type, the AI answers Open questions, exploration, help Burying actions in text; long answers; no structure
Generative UI The AI composes interface pieces on the fly: a table, a chart, a form, a card Showing results in the right shape Inconsistency with the design system; layouts nobody tested
Embedded AI AI acts inside the existing screens: fills a field, adds a row, proposes a change Expert, repeated work Changes people don't notice; no undo

The best products mix them: a short conversation to state the goal, generative pieces to show results, and embedded changes the user can see and undo. See Generative UI and Victor Yocco on matching the AI's form to what people want to do (Smashing, Jul 2026).


2. The central idea: notices become states

When an agent works inside a product, it sends a stream of small notices (developers call them events): I started; I'm on step 2; I changed this value; I want to use this tool; I need your approval; something went wrong. The interface listens and shows each one.

NoticeStarted

The agent began working.

State: a step indicator appears
NoticeStep

It moved to the next step.

State: the current step is highlighted
NoticeData changed

It changed a value or a record.

State: the field is marked, with undo
NoticeUsing a tool

It is searching, checking or calling a service.

State: "Checking stock…" where it matters
NoticeNeeds approval

It is waiting for a decision.

State: a confirmation dialog
NoticeError

Something went wrong.

State: the problem shown at the field, with a fix
Each notice an agent sends is an interface state a designer has to design, like empty, loading and error states. Diagram: Marcus Andrey, CC BY 4.0.
Two red-and-white railway semaphore signal arms with red lamps, mounted on a gantry in front of an ivy-covered brick wall.
Railway semaphores told drivers what the system was doing long before screens existed: one arm, one state, readable at a glance. An agent's notices need the same clarity. Photo: Jeremy Segrott, cropped, Wikimedia Commons, CC BY 2.0.

For designers this means: the list of notices is the list of states to design. Each needs a visual treatment, a place on the screen, words, and an accessible announcement (see Accessibility with and for AI). A notice nobody designed becomes either nothing on screen (the agent seems frozen) or a raw message dumped into a chat.


3. A 2026 example: AG-UI

AG-UI (Agent-User Interaction Protocol) is an open protocol that standardises these notices so any agent can drive any interface (GitHub; docs). As of November 2026:

  • It's open source under the MIT licence, created by the company CopilotKit and developed in the open.
  • Version 1.0 of the specification came out in September 2026; the project changes almost daily.
  • The 1.0 specification defines 31 kinds of notice in eight families: runs and steps, text messages, tool calls, reasoning, state, activity, sub-agents, and pass-through (events).
  • It supports human-in-the-loop directly: a run can stop with an "interrupt", the person approves, edits or rejects, and the next run resumes (interrupts).
  • Agent frameworks from Microsoft, Google, AWS, LangChain and others support it.

See it working in the AG-UI Dojo, a gallery of live demos.

Who talks to whom

MCPAgent ↔ tools and data

Connects an agent to tools, files and services.

Anthropic, now at the Agentic AI Foundation (Linux Foundation)
A2AAgent ↔ agent

Lets agents hand work to each other.

Started by Google, now a Linux Foundation project
AG-UIAgent ↔ interface

Streams the agent's notices to the screen the person uses.

Created by CopilotKit, MIT, spec 1.0 (Sep 2026)
Who talks to whom: one agent often uses all three. The three-part picture is how the AG-UI project describes itself. Diagram: Marcus Andrey, CC BY 4.0.

This three-part picture is how the AG-UI project describes itself; it's a useful map, not an industry standard. MCP and A2A are now hosted by Linux Foundation bodies; AG-UI has no foundation yet.


4. Designing the states: a worked example

The flow (fictional): a person asks an agent to "create an order for three shipments to the Porto warehouse".

Notice What the person should see
Run started A step indicator appears: 1 Collect details · 2 Check stock · 3 Confirm
Step started Step 1 highlighted; the order form opens in the main screen, not in the chat
State changed Fields filled by the AI are marked (a subtle tag "filled by AI") with an undo on each
Tool call "Checking stock…" next to the shipments table, then the table refreshes using the design system's table, not text
Needs approval A confirmation dialog summarising the order, with Edit, Approve and Cancel
Error The problem is shown at the field ("Address not recognised"), with a fix, not a generic chat apology
Run finished A short summary and a link to the activity log

5. Common mistakes

  • Everything happens in the chat; the real screen never changes.
  • The agent works silently for 30 seconds: no "started", no progress.
  • AI-filled values look identical to the user's own input.
  • Approval is a "Yes" typed in a chat instead of a clear dialog.
  • Errors are apologies in prose instead of messages where the problem is.
  • Generated layouts ignore the design system.

6. Vocabulary

Term Meaning
Notice (event) A small message an agent sends to say what it's doing
Generative UI Interface pieces the AI composes on the fly
Protocol An agreed format so different systems can talk to each other
AG-UI An open protocol for notices between agents and interfaces (2026 example)
MCP The protocol that connects agents to tools and data
A2A The protocol that connects agents to other agents
Interrupt A pause where the agent waits for a person's decision

7. Try it (no code needed)

  1. Take the flow in section 4, or one from your product (without confidential names).
  2. List the notices the agent would send, in order.
  3. For each, sketch the state: where on screen, what it looks like, the words, what a screen reader announces.
  4. Check: is there any moment where the person can't tell what the agent is doing? Add a state for it.

Reflection questions: Which of your product's flows would work better embedded in the screen than in a chat? Which notice would your users miss most if it were silent?


8. If you lead a team

  • Ask engineering for the list of notices the agent sends, and make "one designed state per notice" part of the definition of done.
  • Add agent states to the design system: step indicator, AI-filled field, tool activity, approval dialog, inline error.
  • Run a design review on the silent moments: anywhere the agent works without telling.

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.

Agentic design · Designing for uncertainty · Generative UI · MCP · Streaming responses · AI error states · Accessibility with and for AI

Saved only on this device.

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