On context management in long-running agents