AI agents for image generation can now get better at their own jobs with each run.
Researchers introduced COMFYCLAW, a framework built around ComfyUI — a node-based image generation tool — that treats workflow construction as typed graph editing. When an agent makes an invalid edit, the system automatically rolls it back. A vision-language model inspects the output, spots visual failures, and converts them into repair instructions the agent can act on. The key mechanism: instead of discarding that feedback, COMFYCLAW distills it into a reusable skill library that grows over time and gets shared across future runs.
Most agent frameworks treat each run as a blank slate, which means the same mistakes get repeated. COMFYCLAW's bet is that logged trajectories, errors, and verifier feedback are training signal, not noise — and that accumulating them into callable skills compounds over time. Across four benchmark splits and three agent models, it outperformed a verifier-only baseline that lacked skill evolution, and human annotators preferred its outputs.
ComfyUI already has a large hobbyist and professional user base, so a framework that makes agents more reliable there has a ready audience. The harder question is whether skill libraries that evolve inside one workflow tool transfer anywhere else — or whether this is a well-executed solution to a deliberately narrow problem.