I am looking forward to models that have post-training on selectively compacting their own context - so they can choose what to forget and what to edit. That alone would make coding agents multiple times more powerful on large code bases.