Agent harnesses repeatedly invoke reasoning after tool feedback, even when an earlier call has already fixed the next action. We introduce HarnessMM (Harness…
机构:清华
来源:arXiv 2610.32813 | AI4Papers 论文推荐平台