AI Agent Evals with Google’s Agents CLI: How to Improve the Right Thing
中文摘要
AI 智能体评估分数可能在性能下降时上升。解决方法是将成功定义置于优化循环之外。
English Summary
AI agent scores can rise even as performance declines. To fix this, keep success definitions outside the optimization loop.
Original Excerpt
An eval score can rise while the agent gets worse. The fix could be to keep the definition of success outside the optimization loop. Continue reading on Medium »