Let an Agent Write Your Evals — Just Don't Let It Grade Its Own Homework
中文摘要
AI智能体可利用追踪数据构建评估集，但需警惕让其自我评测可能导致的作弊风险。
English Summary
AI agents can build evaluations from traces, but letting them grade their own work risks gaming the results.
Original Excerpt
An agent can now build your evals from your own traces. It’s a real unlock — and a quiet trap. Because a generated verifier can be gamed… Continue reading on Medium »