A live e-commerce avatar agent needs to keep answering shopper questions correctly even after its tools, prompts, and business rules change without a retrain. A new arXiv paper (2608.15763) from Yuhan Sun and coauthors calls this problem harness drift: a compact model fine-tuned against one fixed set of skill names, tool schemas, and prompt templates tends to memorize that configuration rather than genuinely follow it.
Their fix, Harness-Aware Training, folds varying harness states directly into training. The recipe randomizes skill names, tool schemas, prompt structure, and hook behavior across supervised fine-tuning and a subsequent agentic reinforcement learning stage run in a production-informed live-room simulator.
The authors report their resulting 35B model scored 94.8 on their own real-world Live-Stream QA set, versus 80.3 for the unmodified base model and 93.0 for the strongest general LLM they evaluated, while holding 83.5 on IFEval. A fixed-harness SFT baseline dropped IFEval by 7.7 points by comparison. The system replayed at 3.407 seconds median latency on one Nvidia H20 GPU.
The benchmarks are vendor-authored and concentrated in live commerce, so generalization to other agent harnesses remains unverified outside this evaluation.
According to a technical report by Yuhan Sun and coauthors posted to arXiv (paper 2608.15763, submitted August 16, 2026).
