AgentSociety

This paper sources its logic from a lineage of computational social science—but reborn in LLM form. It's less Durkheim, more Turing+SimCity+Rawls. The "heritage" here is simulated agency; the "source" is synthetic humanhood at scale.

10,000 LLM agents interact within a shared world, producing 5 million+ simulated behaviors. Commons are simulated as civic structure, economics, and news environments. Data is both emergent and engineered—a living corpus.

The trade-off is epistemological: control vs. authenticity. These agents are puppets and prophets. There's no ethnographer here—just code as covenant. Sociologists risk Faustian blindness if they forget these agents were born from our models.

Each agent has traits, goals, memory, and evolving identities. This is front-facing identity as simulation—AI not as assistant, but as a mirror of social performance and encoded personality. Identity here isn't assumed—it's enacted through interaction.

Can you trust synthetic society? The paper tests for realism—comparing outputs to real-world studies. It's reflexive posture, looping back to validate, authenticate, and re-orient design. Trust becomes a method, not a default.