📄 中文摘要
在LLM代理中，工具链的调用可能导致级联失败问题。初始调用的输出可能存在缺陷，后续调用会接受并误解这些输出，最终导致整个链条失效而未抛出错误。这种错误传播是构建可靠LLM代理的主要障碍。研究表明，早期错误的传播会导致后续失败，影响系统的稳定性。调试这些失败的过程揭示了工具链的脆弱性及其实际失败模式，同时也指出了一些在生产环境中有效的模式。理解这些问题对于提高LLM代理的可靠性至关重要。
LLM中的工具链是什么？为何会失败以及如何思考编排
📄 English Summary
What Is Tool Chaining in LLMs? Why It Breaks and How to Think About Orchestration
Tool chaining in LLM agents can lead to cascading failure issues. An initial call may produce slightly malformed output, which subsequent calls accept and misinterpret, ultimately causing the entire chain to fail without throwing an error. This error propagation is a significant barrier to building dependable LLM agents. Research indicates that early mistakes can cascade into later failures, impacting system stability. Debugging these failures reveals the fragility of tool chaining and its actual failure modes, while also highlighting effective patterns in production environments. Understanding these issues is crucial for enhancing the reliability of LLM agents.