Players are queried in order for message actions. This can lead to weird ongoing conversations such as 5 rounds of all players agreeing that they should go somewhere before actually going somewhere.
With gpt models, the conversations rarely feel super natural. It is hard to create conflict between the simulated agents.
Simulating a few rounds can definitely take a while.
For example, gpt4o-mini likes to make it's agents play a round of trivia. It always answers all trivia questions correct with all characters.
I guess that does make them more agentic somewhat ^^. E.g. I wanted to make the agents argue about a care crash, but they decided to just hang out and play taboo or trivia games instead.