The paper introduces COMFYCLAW, a system that helps AI agents build and refine image generation workflows in ComfyUI. It treats workflow construction as editing a typed graph, with tools organized by stage and automatic rollback of invalid edits. A region-level verifier detects visual failures and suggests repairs.
The key innovation is a self-evolving skill library: trajectories, errors, and verifier feedback from previous runs are distilled into reusable . Across multiple benchmarks and models, COMFYCLAW outperforms systems without skill evolution, and human annotators prefer its outputs. The work shows that evolving a library of reusable skills from past experience improves agent performance on recurring visual workflows.