Compared to Meta Harness, will the performance gain be more significant?

[beep-bebop]

Excellent work! Looking forward to related experiments~

[Yif-Yang]

@beep-bebop Thanks for your interest! The focus of this work is skill optimization. Skills are easy to deploy, transferable across models, and well suited to serve as a learnable adaptation layer for agents.

We are not focusing on modifying harness code to build a better end-to-end harness. Instead, SkillOpt is designed to work with different harnesses and models. In that sense, it does not conflict with Meta-Harness; I would view them as complementary.

A more detailed comparison is definitely something we will include, and we are already working on it. Please stay tuned for our next work, which we expect to release within about two months. It will discuss Meta-Harness and other harness-code-modification-style approaches in more detail.