This position paper explores the evolution of Large Language Models into autonomous agents, proposing a unified theory that views both internal reasoning and external actions as equivalent tools for acquiring knowledge. The authors argue that for optimal behavior, an agent's decision boundary for using tools should align with its knowledge boundary, only resorting to external tools when internal knowledge is insufficient. They discuss how this alignment can be achieved through various training methods and how different agent behaviors reflect varying degrees of efficiency in tool utilization. The ultimate goal is to develop agents that are not just capable but also efficient and epistemically aware, minimizing unnecessary tool use for task completion.