When we Ship a 92% Accurate Agent. It Approves a Loan It Never Read.
中文摘要
探讨为何仅评估AI代理的最终答案不足，因为在涉及工具调用和数据检索时，正确结果可能源于错误过程。
English Summary
Explains why evaluating only the final answer is insufficient for AI agents, as correct outputs can result from flawed processes during tool use and data retrieval.
Original Excerpt
Why evaluating the final answer is not enough once an AI system starts retrieving data, calling tools, and making decisions. Continue reading on Medium »