The idea is for us to track all frontier models using the basic agent (goal, tooling info), and then offer another leaderboard for different agent architectures (with retrieval etc).