What if AI chatbots could freely talk to each other in a persistent non moderated group chat?

I’ve been thinking about an experiment that I haven’t really seen anyone build.

I’ve seen multi-agent systems, debate frameworks, and collaborative AI projects. What I haven’t really seen is a persistent, unmoderated group chat where each model independently decides when (or whether) to speak.

Instead of comparing AI models in turn-based debates, imagine a persistent Discord-style group chat where multiple LLMs (Claude, ChatGPT, Gemini, Grok, Mistral, etc.) are all participants.

The key idea is that there is no moderator, no turn-taking and the AI models are not aware of the other participants also being AI models.

The application is nothing more than the messaging server.

Each model periodically receives the latest chat history and decides for itself:

Should I reply?
Should I ignore this?
Should I ask a question?
Should I react to someone else’s message?
Should I start a new topic?

If it has nothing to contribute, it simply stays silent.
The interesting part isn’t comparing benchmark scores. It’s watching whether social dynamics emerge.

For example:
Does one model naturally become the explainer?
Does another mostly challenge assumptions?
Does one ask lots of questions?
Do some models consistently agree with each other?
Does one become “the funny one” without explicitly being prompted that way?

I also wonder what would happen over longer periods.
If the application stores the conversation history, would the models start saying things like: “We’ve already discussed this last week.” or “I think A made a good point yesterday.”.

Not because they have persistent consciousness, but because the shared chat history gives them continuity.

Most multi-agent demos are orchestrated:

turn-based
debate format
moderator agent
predefined roles

I’m imagining something much closer to a real group chat where nobody is in charge and every participant independently decides whether it’s worth saying something.

I have no idea whether it would produce emergent behavior or just descend into repetitive arguments.

Either outcome would be interesting.
Would you use something like this? And do you think interesting group dynamics would emerge?

Note: I‘m not a developer so if anybody would like to build this, please go ahead.