The interesting claim in AutoDesign isn’t the poster-generation score. It’s that the harness, not the model, became the thing being optimized. A meta-harness optimizer watches rollout feedback and drives a code agent to rewrite the scaffolding around a fixed model, and that scaffolding keeps getting better on its own.
The numbers make the point concrete. Across seven code-agent-model configurations, dropping in the learned DesignHarness lifted the average benchmark score from 54.99 to 67.39 — same models, better harness. On the main track it hits 78.32, past a closed commercial design system by roughly seven points, inside a 40-minute autonomous loop of 253 tool calls and 11 editing turns for under three dollars. The harness carried most of that lift.
This is worth sitting with if you run agents in production. We pour effort into prompt tuning and model selection, then treat the harness — retries, tool routing, fallbacks, when to stop editing — as static plumbing you hand-tune once and forget. AutoDesign treats it as a learnable artifact that accumulates reusable experience, which is a different unit of engineering than a prompt. The HF paper page frames the whole thing as aligning with human design priors, and that’s the part I’d watch closely: recursive self-improvement only stays useful while those priors hold and the eval that scores each rollout measures what you actually care about. Point a self-rewriting harness at a weak reward and it will optimize the weakness beautifully.
So here’s the real question for anyone maintaining an agent stack: if your harness could rewrite itself against your current eval overnight, would you trust the result — or would you first go fix your eval?