Agent efficiency is not just about choosing a smaller model or writing a shorter prompt. In a tool-using agent, cost compounds across the entire trajectory: every extra step replays context, every cache miss turns discounted input into full-price input, and every oversized tool result is replayed until the next user request.