Seonglae ChoMeta-Harness Self-Improving AI
Meta-Harness is an outer-loop system that automatically searches over harness code. Its core objective is : for a fixed LLM and a task distribution , it seeks the harness that maximizes the expected reward of trajectories . When multiple objectives (e.g., accuracy and context cost) are involved, evaluation is done via Pareto dominance.
In each iteration, what changes is not the prompt but the entire agent system code → including tool policy, loop logic, error handling, and context strategy.
Meta-Harness
Meta-Harness automatically optimizes model harnesses — the code determining what to store, retrieve, and present to an LLM — surpassing hand-designed systems on text classification, math reasoning, and agentic coding.
https://yoonholee.com/meta-harness/
