A newly posted research paper examines how one AI model can improve another's performance not by changing its weights, but by designing the environment in which it operates. The work, titled Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI, frames the problem as test-time AI-for-AI: a "Builder" learns to construct better execution environments, or harnesses, for a "Target" while both models remain frozen.
To make the Builder's experience reusable, the authors introduce the concept of a Meta-Skill — principles that specify when support is needed and what resources should be provided. The Builder derives these principles from the Target's execution feedback on a development set, then applies the resulting frozen skill bank to construct harnesses for tasks it has not seen before.
The approach was evaluated on Harness-Bench and NewtonBench. According to the paper, using the full bank of meta-skills improved macro-average performance by 8.95 percentage points compared with no-skill construction, and by 12.02 points compared with delivering the same skill bank directly to the Target.
The authors argue these results highlight the value of translating experience into executable support rather than simply passing information along. They also note that gains observed when the same model serves as both Builder and Target suggest a possible route toward system-level self-improvement through learning to build better environments.
The paper is available on arXiv, with accompanying code hosted on GitHub.