One thing I do find is that subagents are helpful for performance -- offloading tasks to smaller models (gpt-oss specifically for me) gets data to the bigger model quicker.