This is more serious than most people think Every agent-configuration rule written in the last year assumed the same failure mode, the model runs too far. ask before every action. read every doc first. stop and wait for review after one change. astra doesn't have that failure mode. it stops, sometimes before real work gets done, and the same rules that used to protect you are now what's blocking it. the fix is documented directly by openai's own team. full masterclass below.