JIT-Agent is a harness intelligence model that generates task-adaptive agent harnesses on the fly for arbitrary off-the-shelf LLMs, formalizing the harness — memory management, planning, action protocol, and tool orchestration — as a composable, machine-generatable artifact. Equipped with JIT-Agent, DeepSeek-V4-Flash surpasses GPT-5.6 on the DeepSearchQA and OdysseyBench benchmarks, and the resulting harnesses are performance-competitive with mature hand-built agent runtimes such as OpenCode and Claude Code. The authors describe it as the first model purpose-built for just-in-time harness generation, framing harness design as a trainable and transferable capability separate from the base model.
