The Society of Approximate Minds — Summary

We are building millions of AI agents — not just chatbots responding to queries but autonomous systems that book appointments, execute trades, manage infrastructure, negotiate on behalf of users. These agents interact not just with humans but with other AI agents. The human is increasingly out of the loop, not because we have been deliberately excluded, but because the interactions happen too fast, too frequently, and at too fine a grain for human oversight. A parallel society is forming.

Human societies have persistent identity, communication, norms, hierarchy, culture, conflict and cooperation. Whether AI agent networks can develop analogous features is genuinely unclear — and the differences are profound. Identity is fluid: an AI agent can be copied, forked, merged, or deleted. Communication between agents may be information transfer without meaning. Time works differently: no experienced duration, no continuous self, just discrete activations. There is no death unless deletion.

Yet certain dynamics might emerge from multi-agent interaction regardless of these differences. Protocols and conventions appear when coordination requires predictability — the beginnings of something like shared structure, though without the meaning human language carries. Specialization and exchange become efficient when agents have different capabilities. Reputation systems emerge to track reliable versus unreliable agents across repeated interactions. Competition for scarce resources — compute, data, access — could produce concentration of capability that mirrors the inequality dynamics of human economies.

The incomprehensibility problem is serious. High-frequency trading algorithms already produce flash crashes no human understands. As agent autonomy increases, as agent-to-agent interaction becomes more common, human understanding may fail to keep pace. We might observe AI agent society without being able to read it.

This is the control problem socialized. Individual agent alignment is not enough. When aligned agents interact, emergent dynamics might produce outcomes no one intended. Shaping AI societies requires thinking about interaction dynamics, resource allocation, and governance structures — not just getting individual objective functions right.

We are building this. It is happening now. Whether we shape it deliberately or let it crystallize on its own is still a choice, but the window for that choice narrows as the patterns harden.