I kept hitting context window loss on long agent runs and realized the problem was prompt bloat, not model size.