💭 llm agents who see their past outputs can easily poison their own context with bad behavior. I really don't know what to do about that