Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when…
机构:Microsoft
来源:arXiv 2610.26632 | AI4Papers 论文推荐平台