Where this gets really complicated is when you are chaining many LLM calls together (basically any agent). A slight deviation in the call stack can throw off everything else.