AutoDesign's Learned Harness Beats Claude Design by 7.45 Points and Runs a 40-Minute Autonomous Loop for Under $3

arXiv 2608.13560·high signalXBlueskyLinkedInCopy link

AutoDesign puts a meta-harness optimizer above a code agent, letting it recursively rewrite its own harness from rollout feedback rather than treating the scaffold as fixed. On the new PosterBench Main Track — 100 papers across five disciplines for paper-to-poster generation — it scored 78.32, beating the closed commercial Claude Design system by 7.45 points, and dropping the learned DesignHarness into seven other code-agent configurations raised the average score from 54.99 to 67.39. The economics are the story for solo builders: one fully autonomous run executed 253 tool calls and 11 editing turns in 40 minutes for under $3, reaching average conference-poster quality in blind human evaluation.