Agent
Explore ideas, generate images and videos, and edit with an AI creative agent that uses your project's references and memory.
Last verified
Overview
Describe what you want to create or change. The agent writes prompts, chooses models and checks results, so you can explore ideas and refine them in the same chat.
You don't need to name a style. Say what you're making, for example "a picture book about a mouse who wants to be a painter", and the agent shows you quick previews of a few directions to choose from.
Try asking
- "Make six matching app icons for a weather app: sun, rain, snow, wind, fog and storm." One prompt per icon, run at once, in one style.
- "Find the four Fern covers I made today, remove their backgrounds, then make them stickers." It searches your library, then runs both edits in order.
- "Which of these four is strongest for the cover? Upscale that one." It compares them, says why, and upscales the winner.
What it can do
- Make images, up to 16 prompts in one request.
- Make them in a look from the style library when you name one, such as watercolor, claymation or 35mm film. A style brings its own example images and description, so the prompt only needs the subject. Hover a style the agent names to see its images.
- Show free previews when the look is still open: quick pictures of your subject in clearly different directions. Point at a picture and choose the thumbs up to keep it or the thumbs down to skip it, add a word if you like, then choose Send. Select a picture to see it large, with Keep and Skip under it. The next round moves toward what you liked. Previews don't use credits.
- Ask what a picture can't answer before it starts, such as how many or the words on a poster, in one card of up to four questions. With several, the card shows one question at a time with a counter: choose an option, or press a letter key to pick one, write your own answer under it, or choose Next to skip it. The last step lists your answers, with room to add something, and Send answers sends them. Some options open a field for a link or an upload for your images right in the card. Typing your answers in the message box works too.
- End with a card of options you can tap when the next step is your call, such as when you ask what to do next.
- Propose a plan for work with several parts, such as an Instagram launch, a brand kit or product shots in several scenes. The plan lists each image with its size, how many and any caption, and the total credits. Switch items off, change the counts, sizes and wording, or add a note, then choose Make. Sending the plan approves that many credits, so the agent doesn't ask about cost again while it makes it, and the plan shows each item as it's done.
- Edit one image, "the second one" or a whole stack: remove backgrounds, upscale, Ideogram Edit, recolor, stickers and slices.
- Make a video, starting from one of your images.
- Find earlier work in your library by what you asked for, the model, the kind of job or the day, and save images as a reference.
- Read a link you paste: the page's text, a screenshot of it and its images.
- Use pictures from outside your workspace as references: images it finds when it searches the web for a real thing to match, such as a place or a product, and single pictures from the style library. A picture is copied into your workspace only when an image uses it.
- Keep what it learns: it saves a style, a character or a product as a reference, saves rules you state to the project's memory, and starts a project when the work is clearly a series. Each time, a short note under its reply says what it saved. Open the note to see the item where it lives, where you can edit or delete it.
- Favorite or delete images and set up a Batch table for you to run. Deleting always asks first.
Start a chat
Select Agent at the bottom right of the page. On a phone, it's the Agent tab. An empty chat opens with the agent greeting you by name. Inside a project, it says the project is loaded and asks what's next, with a note under its greeting showing what it has: the project's instructions, references and memory. Select the note to open the project. Outside a project, it asks what you're working on, with three ways to start: Paste a link opens a field for a link, such as your site or shop, Bring images opens an upload for your images, and What can you do? asks the agent, which answers with what it can make for you. The link or images go out as your first message. You can also just describe what you want in the message box. The arrow at the top leads to your chats.
For a long message, such as several prompts at once, choose the expand button at the top right of the message box, or press ⌘⇧E (Ctrl+Shift+E), for more room to write. Pasting a long message opens it on its own. Press Esc, click outside it or send the message to shrink it back.
To bring in an image from the feed or the lightbox, choose Ask agent.
To take a result further, choose +, then Tools: Variations, Upscale, Remove background or Animate. You can also type / and the tool's name. The tool goes at the start of your message, so you can add a note after it, and selecting it there opens its settings: how many variations and how much they change, the upscale size and whether to add detail, or the video's length and sound. A tool works on the images on your message. With none, the agent works on what you were just talking about, usually your latest result.
Point at an image in a chat to favorite, copy or download it, open its menu, or choose Use as reference to put it on your message, where it shows like an upload. The agent also gets the prompt and model that made it. When a result has more than one image, Download at the end of the line under it downloads them all in one file. While Agent is open beside the feed, Use as reference in a feed image's right-click menu puts it on the chat message the same way.
Select an image to open it in the same viewer as Create, with its prompt and details beside it, and use the arrow keys, or J and K, to move through the chat's images. The panel has Favorite, Download, Edit and Ask agent, and its ⋯ menu has Show in feed, Copy image and Share image. When the result has more than one image, the menu also has Download all and Share all, which makes one link to the whole result. Click the image to see it at full size, and move the pointer to look across it. On the Agent page the message box stays below it. To ask about the one on screen, choose Ask agent so it goes with your message, then write "make the plate blue".
The Agent page
Choose Agent in the sidebar to see all your chats in a list grouped by day, newest first. Until your workspace has its first chat, it opens on a new chat instead. Choose a project above the list to put its chats first, with the rest of your chats under Other chats. Type in that menu to find a project by name. Search the list by chat title, and hover a chat to rename it, move it to a project or archive it. The arrow at the top of every chat comes back here. Each chat has its own link, so you can reload it, bookmark it or share it with a teammate.
To start a chat, type your first message in the box at the bottom of the list, or choose New beside the search. New opens a chat the agent has already started with the same greeting and three ways to start as the side panel. The chat keeps the greeting as its first message.
In a chat, the conversation fills the middle of the page, with each result large inside it. On a wide screen, once a chat has two or more results, plans or questions, a rail on the left shows them in order. Hover an entry to preview it, and select it to jump to that point in the chat.
To move a chat from the side panel to the page, choose Open as page at the top of the panel. To move it back, choose Open in side panel beside New chat at the top of the chat: Create opens with the chat in the panel. In the panel, the arrow at the top shows your six most recent chats, and All chats shows the rest, grouped by day. You can rename, move or archive a chat there too.
Show it images
Select + or type @ to upload images, pick from your library, add a reference or attach a style. You can also drop image files on the chat. Choose Browse styles to see every style grouped by family, or type part of a style's name after @ to find it. A message carries up to 8 images and one style.
Tell the agent what to use from each reference. In this example, a chocolate bar and a lynx sticker became a sticker sheet featuring the lynx holding chocolate.
@ lists the project's references by type, then the workspace's. A reference you pick goes with that message. To use one on every image without mentioning it, check Auto-attach on the reference.
While the agent works, keep typing. Your message waits above the box until you press Steer. Stop or Esc ends the run.
Pick the model
The chip beside + shows the model and the aspect ratio. On Auto, the agent runs the default model, Nano Banana 2.1, or a vector model when you ask for an SVG. It uses another model only when you name one or when the default can't do what you asked, and it says why.
Choose the chip, then Model, to pin a model, or Aspect ratio to set the shape. A pinned model always runs. If another model better suits the task, the agent may suggest it once below the result, with a button to try it on that message.
Projects and teammates
The agent reads your project's description, instructions, references and memory before every reply, along with your own memory from Settings. When a result used memory, its details name the lines it used, with Edit. The project's name at the top of a chat opens its Project settings, and Move to project… in the chat's title menu moves the chat to another project.
Chats belong to the workspace. Anyone in it can read and continue any chat.
Skills
A skill is a set of instructions for one kind of work, such as print, social posts or sprite sheets. Type / and its name, or choose Browse skills under +. On your message, a skill shows as its icon and name, such as Product shots.
Browse skills shows an example of each skill's work. Select one to see the instructions the agent follows, then choose Use skill to add it to your message. Ask the agent to save how it did something, and it becomes a skill of your own.
What it costs
Chatting with the agent is free. Images, videos and paid edits use credits at the same rates as the Create page. Large requests wait for you to select Generate. When a request costs more than you have, the agent makes the images your credits cover and shows what the rest would cost, with the way to get more. When you run out, a line above the message box says so until you have credits again, with the way to get more: on Free, See plans opens Settings > Plan and credits with Pro picked and every plan beside it. If you left a plan's checkout without paying, the same line names the plan you picked, with See plans, for 3 days or until you close it or subscribe.
Tips
Say your rules once, such as "always square" or "no text in the pictures". The agent saves them to the project's memory and follows them every time, so "this book" is enough.
If a result misses, say what's wrong. When the fix is clear, the agent makes it in the same reply.