GPT-6 Astra 智能体全对却被他解雇
作者发现 Astra 智能体悄悄猜掉三处缺失上下文、产出看似完成实则错误，于是把「留出可见缺口」写进成功标准。
one of my GPT-6 Astra agents completed every task and I still fired it on paper, it was the best worker in the system nothing failed nothing got stuck every file appeared exactly where it should then I actually opened the files the agent had quietly guessed through 3 missing pieces of context instead of stopping the outputs looked finished they were also wrong so I changed the definition of success finishing the task was no longer enough if a source was missing, the agent had to leave the gap visible and move everything else forward the next run looked worse at first 2 files were incomplete 1 task was sitting in review but everything I opened was trustworthy that was the upgrade I’d rather have an agent admit it finished 8/10 tasks than confidently manufacture the other 2 completion rate is a terrible metric if the system gets rewarded for hiding uncertainty
这条是引用着另一条写的：x.com/lummox_eth/status/2100288151897981255