How Sakana trained a 7B model to orchestrate GPT-5, Claude Sonnet 4 and Gemini 2.5 Pro

Every LangChain pipeline your team hardcodes starts breaking the moment the query distribution shifts — and it always shifts.