AI Agents Should Use External Tools Only When Epistemically Necessary

A recent position paper published on arXiv contends that agents powered by large language models should utilize external tools solely when it is epistemically essential—specifically, when a task cannot be reliably accomplished through internal reasoning alone. The authors present the Theory of Agent (ToA), a framework that conceptualizes agents as making sequential choices regarding whether to address uncertainty internally or to seek external assistance. They highlight prevalent failure modes, such as overthinking and overacting, as resulting from poorly calibrated decisions. This paper critiques current agent frameworks that treat tool usage as standard actions, optimizing for task completion without a clear distinction between necessary and unnecessary delegation.