Maybe a similar idea would be running those bots as evaluation tasks. I suppose voting would be a decent proxy objective when you are dealing with very qualitat

Maybe a similar idea would be running those bots as evaluation tasks. I suppose voting would be a decent proxy objective when you are dealing with very qualitative results and don't want to spend a ton of time evaluating them yourself. Spin up a bunch of bots with different ideas and see which ones score the best.