Research
JIT-Agent Trains a Model to Generate Agent Harnesses on Demand, Lifting GLM-5.2 by up to 20.2 Points
Zhang et al. treat the agent harness (memory, planning, action protocol, tool orchestration) as a machine-generatable artifact under a fixed four-module protocol, and train a model to synthesize, repair, and self-evolve harnesses for arbitrary off-the-shelf agentic LLMs. With JIT-Agent attached, DeepSeek-V4-Flash passes GPT-5.6 on DeepSearchQA (+9.1) and OdysseyBench (+4.3), and GLM-5.2 gains up to 20.2 points. The authors report the generated harnesses are performance-competitive with mature runtimes including OpenCode and Claude Code, the sharpest version yet of the harness-over-model argument that has dominated this beat for a week.
↳ Follow the thread