📄 中文摘要
在生产环境中运行 AI 代理时，发现危险的故障并非随机，而是重复出现的模式。这些模式包括相似的幻觉结构、重复的工具调用错误、提示注入变体以及上下文泄漏模式。大多数工具仅提供日志，少数提供追踪，而能够提供结构化故障记忆的工具则寥寥无几。提出了一种模型，其中每个故障都成为一个规范实体，生成确定性的指纹用于执行，新的执行与历史故障进行匹配，并通过策略引擎将信心映射为允许、警告或阻止的决策。关键在于不修改大语言模型（LLM），也不单靠提示。
构建 AI 代理的故障智能
出处: Building Failure Intelligence for AI Agents
发布: 2026年2月16日
📄 English Summary
Building Failure Intelligence for AI Agents
Running AI agents in production reveals that dangerous failures are not random but rather recurring patterns. These patterns include similar hallucination structures, repeated tool-call mistakes, prompt injection variants, and context leakage patterns. Most tools provide logs, some offer tracing, but few deliver structured failure memory. A model is proposed where every failure becomes a canonical entity, a deterministic fingerprint is generated for executions, new executions are matched against historical failures, and a policy engine maps confidence to allow, warn, or block decisions. The key idea is to avoid modifying the LLM and not to rely solely on prompts.