The Evolution of Tool Use in LLM Agents: From Single-Tool Call to Multi-Tool Orchestration

A research team of 14 authors, led by Haoyuan Xu, has published a seminal review paper on arXiv titled 'The Evolution of Tool Use in LLM Agents.' The work provides a comprehensive analysis of the rapid development in the field of AI agents, charting the critical shift from early research focused on whether a model could correctly execute a single tool call (like a calculator or web search) to the current frontier: orchestrating multiple tools over long, complex trajectories. This involves managing intermediate states, execution feedback, and dynamic environments while adhering to practical constraints like safety, cost, and verifiability.

The paper serves as a crucial framework for the field, unifying task formulations and organizing the sprawling literature around six core research dimensions. These include inference-time planning and execution, training methodologies, safety and control mechanisms, efficiency under resource constraints, capability completeness in open environments, and the design of benchmarks for evaluation. It further grounds the discussion by summarizing representative real-world applications, from automating software engineering tasks and enterprise workflows to interacting with graphical user interfaces and mobile systems.

Finally, the authors outline the major challenges and future directions necessary for building reliable, scalable, and verifiable multi-tool agents. This review is positioned as an essential guide for researchers and practitioners aiming to advance AI agents beyond simple assistants into autonomous systems capable of executing sophisticated, multi-step plans in the real world.