Building the Harness Automatically: Self-Play in Code Distills a Text Harness for Black-Box Optimization

Artificial Intelligence is shifting from prompt-based models toward Agentic Harnessing, which utilizes memory and external tools for complex, real-time tasks. Recent research explores building these harnesses automatically through self-play in code to distill numerical search strategies into text. This approach allows agents to learn via executable practice, significantly improving black-box optimization. Specifically, a distilled 197-word primary Harness A reduced Gemini Flash regret by 48% and enhanced performance across various BBOB benchmarks compared to unaided language models.