📄 中文摘要
许多开发者在演示 LLM 代理时，模型能够完美地选择函数、传递参数并生成回答。然而，当将其部署到包含 50 个真实 API 端点的环境中时，系统往往会出现问题。这一现象在工具使用教程中鲜有提及。尽管关于 LLM 工具使用的研究已经相对成熟，明确了有效与无效的使用方式，但这些发现并未广泛应用于主流的 AI 代理构建指南中。通过对学术文献的深入分析，揭示了当前构建代理时的关键发现及其在生产环境中可能遇到的失败模式。
为什么 LLM 代理在使用工具时会崩溃（以及该如何应对）
出处: Why LLM agents break when you give them tools (and what to do about it)
发布: 2026年3月15日
📄 English Summary
Why LLM agents break when you give them tools (and what to do about it)
Many developers find that their LLM agent demos work flawlessly, with the model selecting the right functions, passing clean arguments, and generating coherent answers. However, when deployed with 50 real API endpoints, the system often fails. This gap is rarely addressed in tool-use tutorials. Although research on LLM tool use is quite advanced, with clear insights into what works and what doesn’t, these findings have not been widely incorporated into the mainstream 'how to build an AI agent' blog posts. A thorough review of the academic literature reveals critical insights for building agents today and highlights potential failure modes that can occur in production environments.