Agent-Society Socialism
From the AI Research Compass wiki
Agent-Society Socialism: Populations of interacting agents will develop norms, division of labor and institutions, and intelligence will increasingly live at the level of society.
Beliefs
Agent-Society Socialism looks past single assistants and small teams to whole populations of agents that trade, cooperate, compete, specialize and govern themselves. Its adherents expect that once agents persist, remember and interact at scale, something like a society will form: conventions for communication, reputations, markets for services, shared commons of tools and knowledge, and collective problem-solving beyond any member’s ability. The “socialism” is about the collective, not economics: the interesting unit of analysis is the social whole.
It sits on the Language side because these societies are mostly built from language-model agents talking in natural language, and deep in the Harness half because the substance is the multi-agent system and its institutions. Its paradigm level is neutral: it builds on mainstream language models, even as it imagines a future the field has barely begun to study.
Arguments
- Multi-agent reinforcement learning has shown that competition and cooperation among agents produce autocurricula, with strategies emerging that no designer specified.
- Frameworks for many communicating LLM agents have reported emergent coordination behaviors, such as role differentiation and information sharing.
- Human intelligence itself is heavily social and cumulative; it would be surprising if artificial intelligence scaled purely as isolated individuals.
Critiques
- Many reported “emergent social behaviors” are thin, scripted by prompts, or hard to distinguish from role-play.
- Agent societies are expensive to run and hard to evaluate, with few clear metrics beyond anecdote.
- Safety-minded critics worry that collective dynamics such as collusion and runaway feedback are harder to oversee than any single model.
Representative works
- Emergent Tool Use From Multi-Agent Autocurricula2019
- AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors2023
Relationships
Friends
Frenemies
How to get it
Land on this cell in the full or pro test: x ≈ −0.67, y ≈ −0.90, paradigm ≈ −0.01.