Low Reasoning Effort Is Enough for Routine Office Work by Language-Model Agents

Xiaonan Xu, Wenjing Wu

Purpose: Developers choose how much a language-model agent reasons before it acts. Some pick a high level for fear that a low one is not enough, and pay for it in tokens, time and…