JIT-Agent Makes the Harness Something the Model Writes at Runtime
The harness thread just got its logical next chapter. A National University of Singapore team released JIT-Agent, a model trained to do one meta-level thing: generate the agent harness itself, just in time, customized for the task in front of it. Memory management, planning strategy, action protocol, tool orchestration, all the parts that are currently designed by hand per task, formalized as composable machine-generatable artifacts under a fixed four-module protocol.
The numbers are the argument. With a JIT-generated harness, DeepSeek-V4-Flash beats GPT-5.6 on DeepSearchQA by 9.1 points and on OdysseyBench by 4.3. GLM-5.2 gains up to 20.2 points. Read that again with the week's economics in mind: a cheap open model inside a purpose-built harness outscoring a frontier flagship. It works across DeepSeek, Mimo, and Qwen families, and the system repairs harnesses mid-run and learns from an archive of past configurations, so it improves with use.
The trajectory across the last month is hard to miss. NVIDIA's NOOA showed a handcrafted harness taking Opus 5 from 30 to 100 percent on ARC-AGI-3. Google's EnvHarness had an agent rewrite the environment. JIT-Agent automates the authoring of the harness itself. First we learned harnesses matter more than anyone budgeted for, then we learned environments do too, and now the harness is becoming something models generate for each other. The manual-harness era looks like it will be remembered as a brief artisanal phase.
For anyone building agents, the practical takeaway is that harness design knowledge is compiling into models, the same way prompt engineering compiled into instruction tuning. The durable skill is defining the task and the checks, because the scaffolding in between is becoming a generated artifact. Paper at arxiv.org/abs/2608.25593.
← Back to all articles
The numbers are the argument. With a JIT-generated harness, DeepSeek-V4-Flash beats GPT-5.6 on DeepSearchQA by 9.1 points and on OdysseyBench by 4.3. GLM-5.2 gains up to 20.2 points. Read that again with the week's economics in mind: a cheap open model inside a purpose-built harness outscoring a frontier flagship. It works across DeepSeek, Mimo, and Qwen families, and the system repairs harnesses mid-run and learns from an archive of past configurations, so it improves with use.
The trajectory across the last month is hard to miss. NVIDIA's NOOA showed a handcrafted harness taking Opus 5 from 30 to 100 percent on ARC-AGI-3. Google's EnvHarness had an agent rewrite the environment. JIT-Agent automates the authoring of the harness itself. First we learned harnesses matter more than anyone budgeted for, then we learned environments do too, and now the harness is becoming something models generate for each other. The manual-harness era looks like it will be remembered as a brief artisanal phase.
For anyone building agents, the practical takeaway is that harness design knowledge is compiling into models, the same way prompt engineering compiled into instruction tuning. The durable skill is defining the task and the checks, because the scaffolding in between is becoming a generated artifact. Paper at arxiv.org/abs/2608.25593.
Comments