Comfy Agent

• Building a text-to-image workflow for a product shot from one sentence, then changing the background without touching the model or sampler settings
• Asking the agent to explain an inherited workflow, find where it is slow, and fix the node errors that stop it from running
• Comparing several video models on the same generation by asking the agent to update the workflow with more video models
• Re-running a saved product-shot workflow on a batch of 20 images, or producing ad variants in several aspect ratios
• Saving house rules such as canvas size, background, and validation steps as a personal skill the agent applies to future requests
• Trimming or joining generated video clips and adding an audio track without leaving ComfyUI

Comfy Agent goes after ComfyUI's biggest barrier, which is not image quality but the work of choosing models, wiring nodes, and chasing errors. Because it edits the real graph instead of hiding it, the output stays reproducible and inspectable, which suits people who care about craft and will tune what the agent builds. For now the value is tied to Comfy Cloud: the local version that would bring your own GPU, models, and custom nodes is still on a waitlist, and Cloud usage means paying for agent tokens on top of GPU time. Start in Ask mode on a few real workflows to see how much each session spends before switching to Auto.

About Comfy Agent

ComfyUI is the open-source node-graph tool many creators use for diffusion and video pipelines, and its power comes with a cost: finding the right model, wiring dozens of nodes, and debugging errors takes time. Comfy Agent, launched by Comfy Org on October 1, 2026, is a panel inside ComfyUI that does that work from a chat. You pick a workflow tab in the Agent panel and describe what you want. On a blank tab, it searches for a matching template before building from individual nodes. On an existing workflow, it changes only the part you name, for example adding an upscaling stage or switching to a different model while keeping the sampler settings. It searches the templates, models, and nodes available in your environment, adds, connects, and configures nodes, checks connections, model names, and required inputs before a run, and separates workflow errors from inputs only you can provide. Runs follow a permission you set: Ask needs approval before each run, and Auto lets it run and iterate on its own. It can inspect images, read video and audio properties, do simple media edits such as trimming or joining clips, save personal skills from plain language or a pasted SKILL.md, use Comfy's public skills, and hold up to five chats at once. You can keep editing the canvas while it works, and every result stays an editable workflow with its nodes, seeds, samplers, and models visible, which is the main difference from chat tools that only return a file. The agent runs on Anthropic models (the pricing table lists Opus 4.8), and its tokens are charged in Comfy Credits from the same pool used for generation, after some free initial chats. The tradeoffs: it is Cloud-only for now, so it is limited to Cloud-supported custom nodes and models, even the coming local version sends your messages and workflow information to Comfy's online service, there is no choice of language model yet, cost comes from two meters (agent tokens and GPU runtime), running workflows on Cloud needs a paid plan once the free runs are used, and Comfy still labels the agent as beta.