Figured now was a good time to post this since we recently got surprisingly good results on training an email research agent. Link is above, but will put it here as well since I think it's a good example of RL's promise: https://openpipe.ai/blog/art-e-mail-agent