I've read one-off anecdotes of such rings across conferences. Has there been a study by chairs/organizers about how frequents this is?

In the early days of CS, reviewing loads were typically 20 papers. This required a much greater commitment from the reviewers, but results were much more consistent. For FOCS and STOC, reviewers read EVERY submission

Free food might motivate grad students to do more 😂

The criteria to get accepted to ARR should be similar to TMLR. The presentation slots in conferences and workshops left to PCs to decide. Decouple the two completely.

One of my gripes with ARR (orhers have also raised this) is that reviews and scores are not tied to accept/reject decisions. But coupling them together again brings it closer to the old review system again 😅

An agent would be an LLM to do one specific task (or function, as you refer to it) such as plan (decide to delegate), summarize long context, summarize short fontext

Yes, a single system can delegate to multiple different 'agent' LLMs. So a multi agent system is still just one system but has differing components inside (here, an agent that can summarize novels vs another that can summarize a news article)