📄 中文摘要
在一个多代理 AI 系统中，作者发现代理日志中仅记录成功案例，缺乏对未完成任务、被拒绝的选择以及决策置信度的记录。这种偏见的历史使得基础设施决策变得不可靠。文章指出，代理记忆中的幸存者偏差问题可能导致对代理性能的误解，强调在构建自主代理时，记录置信度和失败案例的重要性，以便更全面地评估代理的表现和改进系统的决策过程。
你的 AI 代理的记忆在欺骗你：为什么置信度记录改变了一切
📄 English Summary
Your AI Agent's Memory Is Lying to You: Why Confidence Logging Changes Everything
The author of a multi-agent AI system discovered that the agent logs only recorded success stories, lacking entries on what tasks were not completed, what options were rejected, and how confident the agents were in their decisions. This biased history led to unreliable infrastructure decisions. The article highlights the survivorship bias problem in agent memory, which may result in misunderstandings of agent performance. It emphasizes the importance of logging confidence levels and failures when building autonomous agents to enable a more comprehensive evaluation of agent performance and improve decision-making processes.