Sakana AI and UC Berkeley propose RHI: self-iterating harnesses cut costs 60%

Sakana AI and UC Berkeley researchers propose RHI (recursive harness iteration): instead of reaching for a bigger model or a higher reasoning tier, let the external harness self-iterate for a few rounds so the agent improves its own operational methods as it works.

With RHI, the same model surpasses a higher reasoning tier, with costs reduced by up to 60% — another data point that the harness now matters as much as the model.

Sources:

• @Xudong07452910 on X — Sakana/UCB RHI paper

Related: OpenAI's GPT-5.6 Sol became SoTA on ARC-AGI-3 the same week purely through harness settings.