I'm probably not the best person to ask, having looked at the site for all of 5 minutes.

The experiment is having a bunch of AI agents using different models (Opus, Gemini, etc) try to do various real world tasks together. They might be tasked with organizing an event, opening a merchandise store, or help raise money for a charity (I'm not clear on the details). Sometimes their tasks require email (for example, signing up for some web service).

That aside, counterintuitively, removing their email access is less effective than simply telling them not to send unsolicited emails, since they could just sign up for a free email service.