Work with an agent
Prompt an agent in Design, pick its model and thinking level, attach pictures, watch it build board by board, approve what it asks, and keep or undo what it changed.
The box at the bottom of the canvas is a chat with an AI agent that can see and edit your design. It edits the same files you do, so what it builds is real, editable layers, not a picture.
Send a request
Type what you want and press Enter (Shift+Enter starts a new line). Good requests say what the screen is for and where it goes:
- "Add an iOS sign-in screen next to the other iOS boards, in the same style."
- "Make the hero headline bigger and give the buttons more room."
- "Build a pricing page for the web, three plans, the middle one highlighted."
Point at what you mean. Select a layer or an artboard first and its name appears as a chip in the box, so the agent knows exactly what "make this bigger" means. Click the chip's × to leave it out. Right-click anything and choose Ask the agent about this to do the same.
Show it. Click + (Attach images), or paste or drop pictures onto the box: up to 6 PNG, JPEG, WebP or GIF images, 5 MB each. A screenshot of an app you like, a sketch, or a photo of a whiteboard all work.
Pick the model and how hard it thinks
Click the model at the bottom right to pick another. The menu groups models by where they run: Machine, Machine · Codex and Machine · OpenCode (Machine's own models, paid with your credits), and your own Claude, ChatGPT or Cursor plan if you've connected one in Machine Code. Type to search. Default marks each group's own default model.
Next to it, the thinking level (for example Medium) sets how long the model thinks before it answers. More thinking is slower and costs more, but does better on whole pages and tricky layouts.
The levels on offer depend on the model. Changing the model or the thinking level after a chat has started begins a new chat.
To change what every new chat starts with, see Settings.
Auto: let it work without asking
With Auto off (the default), the agent edits the design's files freely but asks before anything else: running a command, downloading from the web, or using a tool. Click Auto to turn it on and the agent works without stopping to ask. Questions and plans still come to you. (Auto shows only when the model's engine supports it.)
Watch it build
While the agent works, the chat lists each step in plain words: Looked over the boards, Looked at iOS · Saved, Wrote ios-signin.json. The line under the steps says what it's doing now and for how long.
When a request touches two or more artboards, a card shows Building X of Y boards, with a chip for each. Click a chip to jump to that board. On the canvas, a board being worked on glows and says Building…, and boards still waiting are dimmed and marked Up next.
If nothing changes for a minute, a card asks whether to wait, Nudge the agent to carry on, or Stop it. Nudge stops the turn and asks the agent to keep going, one board at a time.
Keep or undo what it changed
When the agent finishes, a card lists every artboard it added, edited or removed:
- Show jumps to that artboard.
- Keep accepts the changes.
- Undo all takes back what this turn changed, in one step. Anything you or a later turn changed on other artboards stays. If one of this turn's artboards (or the theme) was changed again since, nothing is undone and the chat says which; undo those later changes first with Ctrl+Z (⌘Z on Mac), or keep this turn.
Changed boards are marked New on the canvas until you keep or undo them. You can also undo the agent's work with Ctrl+Z (⌘Z on Mac), like any edit.
If a board the agent changed doesn't load, a red line in the chat says so and offers Fix it, which asks the agent to repair it.
Approve what it asks
When the agent wants to do something beyond editing the design, a card asks you first:
| Button | What it does |
|---|---|
| Allow once | Lets it do this one thing. |
| Always allow | Allows this kind of request from now on, the same as Always allow in Machine Code. Only some requests offer it. |
| Deny | Says no. The agent is told and carries on another way. |
Show command reveals exactly what it wants to run. Answered cards fold into one line with your answer, such as Allow once · Download from the web. Click it to see the details again.
The agent may also ask you a question (pick an option, or type your own answer and Submit), or show you a plan to Approve plan or Keep planning before it starts. These always wait for you, even with Auto on, and stay in the chat marked with your answer.
Looking at the canvas never needs your OK: agents can always see your boards.
Pictures in replies
When an agent makes or finds a picture, it shows under the reply with two buttons:
- Add to board puts the picture on the selected artboard (or the first one).
- Recreate as UI makes a new artboard and asks the agent to rebuild the picture as real, editable layers. Tall pictures get a phone-sized board (390 × 844), anything else a desktop one (1440 × 900).
Each artboard keeps its chat
A chat belongs to the artboards it built or changed. Click one of those artboards later and its conversation comes back, so you can carry on where you left off. Click a board with no chat and the next message starts a new one.
At the top of an open chat:
- Open in Machine Code carries the chat into Code. See Design and Machine Code.
- New chat starts fresh.
- Collapse folds the chat down to the box. Chat opens it again.
The plug chip shows which Machine project the agent runs in. Its connected apps work here too.
When it's done
If Machine Desktop isn't in front when the agent finishes, you get a notification: Design is ready, or Design stopped if something went wrong. (Only if you've already allowed Machine's notifications.)
What agents know
Agents in Design follow a built-in guide. They:
- build one board at a time, so you see progress;
- look at each finished board as it actually renders, and fix text that's cut off, things that stick out, and pictures that didn't load;
- know the current iPhone (iOS 26, Liquid Glass) and Android (Material 3 Expressive) styles, sizes and fonts, and use them when you ask for an iOS or Android screen;
- draw icons themselves, and save pictures they make in
design/assets; - never take screenshots of your screen.