COMFYCLAW: Self-Evolving Skills for ComfyUI Workflows, 2 Jul 2026

COMFYCLAW is an agentic skill evolution harness for controlling ComfyUI workflows, submitted to arXiv on 2 Jul 2026. The paper presents a system that formulates workflow construction as typed graph editing, exposes stage-organized tools, and evolves a reusable skill library distilled from past runs.

What is COMFYCLAW and how does it work?

COMFYCLAW is a framework that treats workflow construction as typed graph editing and equips agents with tools organized by construction stage to build ComfyUI image-generation flows. The system automatically reverts invalid edits, and it uses a region-level vision-language model (VLM) verifier to translate visual failures into actionable repair suggestions; trajectories, execution errors, and verifier feedback are distilled into reusable Agent Skills.

The paper describes several concrete mechanisms: typed graph editing as the construction primitive, a toolset arranged by construction stage for the agent to call, an automatic revert mechanism for invalid graph edits, and a region-level VLM verifier that turns visual failures into repair hints. Those collected signals feed a progressively disclosed skill library that the agent can reuse in later runs.

How was COMFYCLAW evaluated and what were the results?

The authors evaluated COMFYCLAW across four benchmark splits, three agent models, and two image backbones, and report that COMFYCLAW achieves the best average image-generation evaluation score across all six agent configurations. Human annotations in the study also show annotators preferring COMFYCLAW over variants that omit skill evolution.

Evaluation specifics given in the paper include the experimental axes: four benchmark splits, three different agent models, and two image backbones. The paper contrasts COMFYCLAW with a verifier-only baseline, finding that the skill evolution mechanism outperforms that baseline without skill evolution.

Why does this matter?

COMFYCLAW targets recurring, domain-specific workflow construction where memory and reusable skills improve reliability and efficiency. By distilling trajectories, execution errors, and verifier feedback into reusable skills, the framework addresses repeatability and preference retention across runs. For teams building automated image-generation pipelines in ComfyUI, the result is a concrete approach to reduce repeated manual repairs and to improve generation quality via learned workflow patterns.

What to watch

Look for follow-up artifacts the paper links to, such as code or data releases, and for independent reproductions across more agent models and backbones; the paper’s current evaluation covers four benchmark splits, three agent models, and two backbones. Also monitor whether the progressively disclosed skill library approach is applied beyond ComfyUI workflows to other workflow-based agent domains.