Why do AI agents get worse with more context, and how should we manage context windows?