How to Make Tool-Using LLM Agents Efficient? A Survey

Large language models (LLMs) are increasingly deployed as agents that leverage external tools to resolve complex tasks, transcending the limitations of static parametric knowledge. However, tool-using LLM agents often rely on extended context, multi-step reasoning, and frequent tool interactions, incurring substantial computational and financial resources. As a result, efficient tool-using LLM agents that could reduce one or more forms of resource consumption while maintaining comparable task performance have attracted growing research attention. Nevertheless, existing efforts remain fragmented across different stages and resource dimensions. To address this gap, this survey presents a systematic review and a unified taxonomy of efficient tool-using LLM agents. We categorize efficiency techniques along four stages of the agent loop: (i) Efficient Tool Context, (ii) Efficient Tool Calling, (iii) Efficient Tool Reasoning, and (iv) Efficient Tool Execution. By analyzing these methods, we reveal their shared mechanisms and limitations, identify critical open challenges, and highlight promising directions for future research. This survey provides a foundational roadmap to accelerate the development of efficient tool-using LLM agents.