JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution
The paper claims agent gains can come from generating the harness around a model, not just scaling the model itself.
JIT-Agent is trained to synthesize task-specific agent harnesses covering memory, planning, action protocol, and tool orchestration. The authors say it can also repair harnesses during execution and improve by distilling signals from prior harness configurations. With JIT-Agent added, DeepSeek-V4-Flash is reported to beat GPT-5.6 on DeepSearchQA and OdysseyBench, while GLM-5.2 gains as much as 20.2 points. HF Daily Papers' note
JIT-Agent is trained to synthesize task-specific agent harnesses covering memory, planning, action protocol, and tool orchestration. The authors say it can also repair harnesses during execution and improve by distilling signals from prior harness configurations. With JIT-Agent added, DeepSeek-V4-Flash is reported to beat GPT-5.6 on DeepSearchQA and OdysseyBench, while GLM-5.2 gains as much as 20.2 points. HF Daily Papers' note
score 5