What strategies do you use to optimize context window usage and minimize token costs in long-running agent conversations at Adobe?
