Testing Comfy Agents

Building ComfyUI Workflows Through Natural-Language Instructions

Comfy Agent is designed to let users describe an image or video-generation task directly inside ComfyUI, where it can plan, build, edit, validate, and run the required workflow. In the demonstration, the agent creates new workflows, improves prompts, changes image dimensions, works with reference images, and prepares an image-to-video process. It can also search available templates, models, and nodes before adapting a workflow to the resources available in the current ComfyUI environment.

Automated Workflow Editing With Human Oversight

The demonstration shows both the potential and the practical limitations of AI-assisted ComfyUI automation. When a listed VAE is unavailable, the agent can substitute another version so that the workflow can continue, although changing core components may affect the final image quality. Some requested parameter changes also require checking, manual correction, or further experimentation. Comfy Agent can validate node connections, model names, required values, and output nodes before execution, while its permission system allows users to approve each run rather than giving the agent unrestricted control.

Comfy Cloud Availability and the Future of Local AI Workflows

Comfy Agent is currently generally available in Comfy Cloud, with free initial tokens available for new users before usage draws from the existing Comfy Credits system. The service uses frontier models from Anthropic and can inspect images as well as read video and audio properties, giving it broader context than a text-only assistant. Local support for Comfy Desktop is planned, which would eventually allow the agent to work with a user’s own GPU, models, and custom nodes. This points towards a more accessible form of ComfyUI automation, while retaining the editable node graph, workflow transparency, and technical control valued by advanced users.