Tag
This paper explores using general-purpose coding agents to build images through executable programs rather than generative models, introducing the AgentDrawBench benchmark and finding that agents maintain a shared internal visual language-of-thought across drawing interfaces.