JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution

NUS's LV-Lab (first author Guibin Zhang; senior authors Wangchunshu Zhou and Shuicheng Yan) argues the harness — memory management, planning strategy, action protocol, tool/skill orchestration — "can dominate the contribution of the underlying foundation model," yet is designed manually per task. JIT-Agent is a harness intelligence model: it formalizes the harness as a composable, machine-generatable artifact under a fixed four-module protocol and is trained to synthesize a task-adaptive harness on the fly for arbitrary off-the-shelf LLMs, repair it for stable execution, and self-evolve by distilling signals from an archive of prior harness configurations.

A distinct point in the harness-evolution design space: where SkillOpt optimizes a document, AutoSaddler patches a codebase, and WikiSkill grows a knowledge base, JIT-Agent moves harness generation into a trained model invoked per task.