AutoDesign targets systems that turn multimodal content into structured design outputs, arguing that most such pipelines stay static rather than improving from experience. It introduces a meta-harness optimizer that guides a code agent to recursively refine its own harness based on feedback from prior rollouts, demonstrated on the task of generating academic posters from papers. On the PosterBench benchmark, the resulting system scored 78.32 on the Main Track, outperforming a commercial baseline, and integrating the learned design harness improved performance by an average of 12.4 percentage points across configurations.
