You also want prompt/prefix cache. Otherwise there is a lot of duplicate prefill processing if you fork conversation, have a different chat window or anything like that. It therefore also makes subagents much faster.