Agent

The Agent node writes your prompts for you. Connect it to the nodes that make your content, tell it what you want, and every time you hit Run it writes one prompt for each of those nodes: a different prompt for Scene 1 than for Scene 2, and something new next time.

It only writes prompts. It never adds, deletes, or reconnects anything on your canvas.

Inputs and outputs

Inputs (all optional): Text for a brief coming from another node, Images, Video, Styles for a Shots or a Camera Moves node, and a generic Flow input for connecting from earlier nodes. The Video input only shows up when the model you picked can watch video (the Gemini and Qwen models can, and GLM 5.3 Flash), or when a video is already wired in.

Output: a single Prompts handle. Connect it to anything that takes a prompt, such as Image Generation, Video Generation, Voice Generation, or LLM. Each connected node gets its own prompt, written for what that node makes: a composed still for an image, movement and camera for a video, the actual words for text.

Connecting it to a node that does not take a prompt (Compose or Output, for example) costs you nothing. It is skipped.

Writing the brief

Type what you want into the box at the bottom of the card. Every run needs one: your instructions and its memory shape what it writes, but they never decide what this run is about. That stays yours. A Text node wired into the agent's Text input counts as the brief too, so a brief can come from earlier in the workflow.

The card shows how many prompts it will write when you hit Run, then shows each prompt it wrote, labelled with the node it went to.

Naming the nodes it writes for

The agent reads the names on your nodes as intent, so renaming them is the fastest way to steer it. "Scene 1", "Scene 2" and "Scene 3" become consecutive moments in one piece. "Hook" and "Payoff" become two different jobs. "B-Roll" becomes supporting footage rather than the subject.

Leave them unnamed and it treats them as an ordered set and gives them a progression anyway.

Memory

Click the brain icon in the settings bar to open Memory.

Your instructions: style, tone, format, subjects to stick to or stay away from. The agent follows these on every run and writes them into every prompt it produces, so the node downstream hears them too. This is yours: the agent reads it and can never change it. You have up to 20,000 characters, and a counter shows up when you get close. Your instructions save when you click away or close Memory. To remove them, delete the text. If they run over the limit, Memory tells you how much to cut and doesn't save until you do. An AI assistant you connect to your account (see MCP for AI Agents) can also write, change, or delete them for you, and it asks you to confirm before it deletes them.
What it has made here: one line per run, newest first, describing what was distinctive about it. This is how it avoids giving you the same thing twice. You can delete any entry, or clear the lot, but you cannot edit them: they are a record of what happened.

Memory belongs to the workflow, so each workflow builds up its own.

Images and video

Wire images or video in and the agent uses them as reference for what it writes.

It works from your file names and the tags added automatically when you upload, which is usually enough to tell one asset from another and costs you nothing. When an image has no name and no tags, it opens it and looks, up to three times per run. Naming your assets keeps that from happening. An image made earlier in the same run has no name yet, so it is one of the ones it opens.

Video is different: only some models can watch one (the Gemini and Qwen models can, and GLM 5.3 Flash), so the Video input only appears when your chosen model can. If you wire a video and later switch to a model that cannot watch it, the connection stays, the run still works, the agent goes by the file name instead, and the node's warnings tell you so and name what to switch to.

A Gemini model can take up to 20 MB of images and video in one run. If the files you connected are bigger than that together, the node stops before anything is charged and tells you both sizes. Use a shorter or smaller clip, or connect fewer files.

Shots and camera moves

Want a say in how your images and videos are framed and how the camera moves? Add a Shots node, a Camera Moves node, or both, pick what you want on them, and connect them to the agent's Styles input. The one input takes both nodes.

The agent builds your picks into the prompts it writes: shots into image and video prompts, camera moves into video prompts only. A prompt gets one pick from each tab. When you picked several, the agent decides which prompt gets which and which move goes with which shot.

Styles shape the prompts. They are not a brief, so a run still needs one.

Choosing the references for each image

An Image Generation node normally uses every reference picture wired into it. Wire three characters into every shot and all three go into every image, also the ones who are not in that shot.

You can let the agent decide instead. On the Image Generation node, switch on References chosen by Agent (the button with the agent's icon in its settings bar). The agent is then told what is wired into that image, by name, for example "Timon, Pumbaa, Guest, Set A", and picks what the shot needs. The image uses only those.

It is off on every node until you switch it on, and a node with it off works as before.
The agent picks whole connections: a character with all its pictures, or one Router output. It does not pick single pictures out of one.
It can pick none. The image then uses no reference.
The names it chooses from are the ones on your canvas: the name you gave a node, a character's name, the name of a Router output. Clear names make for better picks.
Under each prompt on the agent's card you see what it picked, for example "References: Timon, Set A".
If the agent picks something that is not wired into that image, it gets one more try with the mistake pointed out. If that is wrong too, the run stops at the agent, before any image is made, and the message names the image and the wrong name.
To change a pick, run the agent again. Picks cannot be edited by hand.
A pinned agent keeps the picks that belong to its pinned prompts. If you switch this on for an image the pinned agent already wrote for, it has no pick for that image. The image stops and asks you to unpin the agent and run it again, which rewrites all its prompts.

An AI assistant you connect to your account (see MCP for AI Agents) can switch this on for you, and it sees the picks next to the prompts when it checks a run.

Choosing a model

Click the model picker in the settings bar. The default is GPT-5.6 Luna. The agent runs on Atlas Cloud. Some models read text only; with one of those, the agent uses only your file names and tags, so an image with neither is left unseen, and the run's warnings say so.

Running the node

Select the node to reveal its Run button. Running the whole workflow writes the prompts and passes each one straight to the node it was written for, so everything downstream starts together.

Testing just this node writes prompts and shows them to you without adding anything to its memory, so you can try a brief as many times as you like without using up its sense of what it has already made.

If a node the agent is connected to is switched off, the agent writes no prompt for it and tells you which one. A line under the node names it, for example "Shot 1 motion was switched off, so the Agent wrote no prompt for it." The same line is in the run's warnings.

Pinning its prompts

When the agent has written prompts you like, pin them. Select the node and click the pin next to its Run button. It turns orange, and every workflow run you start then sends those same prompts instead of writing new ones. The nodes after the agent are made again from them. Click the pin again to unpin.

A pinned agent does not run, so it adds nothing to that run's cost and nothing to its memory.
If you connect another node to a pinned agent, the agent writes a prompt for that node only. It reads the pinned prompts first so the new one fits with them, and it leaves the pinned ones exactly as they are. A line under the node tells you what happened, for example "Agent is pinned: wrote 1 new prompt, kept 8 pinned. Unpin to rewrite all." The pin then covers the new prompt too.
Writing the missing prompt takes one model call. If you start the run from your Workflows page or through your AI assistant, the pin does not move to cover the new prompt, so the missing prompt is written again on each of those runs. Run the workflow once from the editor and it is kept from then on.
The agent's own Run button writes every prompt again and takes the pin off.
If you choose "Run only this node" on a node the agent has not written a prompt for yet, the run stops and tells you. Choose "Run this node and after" instead, and the agent writes the missing prompt first.
Scheduled runs ignore the pin: the agent writes fresh prompts each time the schedule fires.

For pins on the other nodes, see Pinning the result you want.

Tips

If a run comes back with something missing, click the Inputs / Outputs button (the eye icon) to see exactly what it wrote and any warnings it left.
An agent connected to nothing that takes a prompt does not run a model at all, so an unfinished canvas costs you nothing.
Nothing you write in your instructions can change what the agent is allowed to do. It writes prompts, and only prompts.