- Difficulty is not very important as criteria here
- What we need is something that shows which is good for which architecture.
- What can matter is having different style of agents
- Token budgets: should we make sure they are all equivilant? or is it a quality metric