In a new research paper, Hugging Face presents JIT-Agent, a trainable model designed to synthesize adaptive agent harnesses for off-the-shelf large language models (LLMs). The work challenges the assumption that agent capability is determined solely by the underlying model, arguing that the harness—encompassing memory management, planning strategy, action protocol, and tool orchestration—often contributes more to overall performance than the foundation model itself.
Traditionally, harness design has been manual, task-specific, and difficult to scale. JIT-Agent formalizes the harness as a composable, machine-generatable artifact governed by a fixed four-module protocol. The model is trained to customize harnesses for specific tasks, repair them for stable execution, and self-evolve by distilling performance signals from an expanding archive of prior configurations.
The results are striking: when equipped with JIT-Agent, DeepSeek-V4-Flash surpasses GPT-5.6 on DeepSearchQA (+9.1 points) and OdysseyBench (+4.3 points). The already strong GLM-5.2 gains up to +20.2 points. In controlled evaluations, JIT-Agent-generated harnesses are performance-competitive with mature agent runtimes such as OpenCode and Claude Code, and consistently improve multi-scale model families including DeepSeek V4, Mimo-V2.5, and Qwen3.6.
The authors claim this is the first model purpose-built for just-in-time harness generation, establishing harness intelligence as a trainable, transferable, and compounding dimension of agent capability that is orthogonal to model scaling. The project is open-sourced with code, datasets, and a website for further exploration.