you can’t see any benefits to studying how LLMs behave in long horizon multi agent tasks?