The agent lives inside ComfyUI, knows the latest models and extensions, and works on the same canvas while you edit by hand. Available in Comfy Cloud since 1 October; the desktop version follows in a few weeks. It is a beta.
- Up to 5 parallel chats with full history; your own skills, public or private; it understands the visual material you hand it.
- Every node stays open. You open it and check every parameter, as before.
- Free to try, drawing on the Comfy Credits you already have. Open source for the agent is not announced.
A ComfyUI workflow looks like the stage plan of a big event: two hundred cables, each in its place, and one that nobody remembers plugging in. That is exactly what made ComfyUI strong, and exactly what drove people away from it. Now an agent that knows the cables walks into the app.
Here is what sets it apart from the ten other "agents for pictures". It does not take you out of the tool. Every control ComfyUI gives you stays where it was: you open the node, look at the parameter, move something and the agent notices. Two of you on one canvas, Comfy Org writes. You know what you want, it knows which model does it and where it breaks.
Comfy Org puts it in one sentence worth writing down: generation got easy, craft did not. Many ways to get an image, few ways to get exactly this image. The agent is placed on the craft side, and that is the right end.
One reservation. Comfy Org gives one artist and a hundred hours with the agent. A hundred hours is a long time for one character. That is not proof the agent saves time; it is proof that someone lasted a hundred hours with it and came out happy. The number that will tell the truth is a different one: how many of the people who ran from the nodes until yesterday are still there after the first week.
The desktop version comes in a few weeks. Until then you try it in the cloud, with the credits you already have, and with a real job, not their demo. Your own job, with your own material, is where you will see how many of the cables it knows.