That's funny, I made comfy agents and a built-in code editor in comfyui about a year ago. When I brought it to comfy Y'all said it would be to much and you wouldn't be interested. Tell Jacob hi for me.
I’m glad you did it. Eventually, I would have circled back to building an agent harness for Comfy. But it probably wouldn’t have been nowhere near as good.
Thank you for this. The MCP and everything that you all are doing over there. It’s nice when a company listens to and cares what its client/customers/users want. If only that was the norm again and not a niche.
The ability to inspect every node while the agent edits the canvas is a useful part of this design. For the twenty product variants example, I would keep the product shape and label as acceptance criteria, rather than count every generated image as a finished ad. I represent the Seedance 2.5 website (https://seedance2-5.co/); that same distinction between generating options and choosing a usable take matters when exploring a product video from a reference image.
Looks good - I use Perplexity Computer to generate and adjust workflows, but it works at the JSON workflow authoring level and doesn't interact with the canvas, so this will be a great step up.
On local, will one need the Windows desktop app or should this also work if you run Comfy locally via a Docker container?
That's funny, I made comfy agents and a built-in code editor in comfyui about a year ago. When I brought it to comfy Y'all said it would be to much and you wouldn't be interested. Tell Jacob hi for me.
I’m glad you did it. Eventually, I would have circled back to building an agent harness for Comfy. But it probably wouldn’t have been nowhere near as good.
Thank you for this. The MCP and everything that you all are doing over there. It’s nice when a company listens to and cares what its client/customers/users want. If only that was the norm again and not a niche.
🫡🖖🏼
🙌🏽🙌🏽🙌🏽👏🏽👏🏽👏🏽
The ability to inspect every node while the agent edits the canvas is a useful part of this design. For the twenty product variants example, I would keep the product shape and label as acceptance criteria, rather than count every generated image as a finished ad. I represent the Seedance 2.5 website (https://seedance2-5.co/); that same distinction between generating options and choosing a usable take matters when exploring a product video from a reference image.
Looks good - I use Perplexity Computer to generate and adjust workflows, but it works at the JSON workflow authoring level and doesn't interact with the canvas, so this will be a great step up.