Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution