Meta-Harness automates harness optimization for LLM systems

Hot take: this is a strong reminder that a lot of LLM performance lives in the surrounding system, not the base model.

• –The core idea is to search over harness code the same way we search over prompts or architectures.
• –The filesystem-backed proposer is the interesting part: it can inspect prior candidates, traces, and scores instead of being limited to compressed feedback.
• –The reported gains are meaningful because they span different task types, including classification, retrieval-augmented reasoning, and agentic coding.
• –The work is most relevant to teams building eval harnesses, agent runtimes, and context-management layers for production LLM systems.