Social theory should be a structural prior for agentic AI: A formal framework for multi-Agent Social Systems

Summary

This position paper argues that agentic AI deployed in social settings has been mis-framed as collections of isolated task-completing agents, ignoring decades of social theory on collective behavior. The authors propose Multi-Agent Social Systems (MASS), a formal triple $(f,g,G)$ — an information exchange function, an influence dynamics function, and an interaction network — and identify four structural priors that any agentic-AI model in a social environment should respect: strategic heterogeneity, network-constrained dependence, co-evolution, and distributional instability. Each prior is given a formal proposition, demonstrated empirically on a real LLM-only social platform, and used to ground a research agenda for modeling, evaluation, and governance. The paper’s core claim is that social theory is not decorative context but a structural prior for agentic AI, distinct from and complementary to alignment work like RLHF.

Key Contributions

• A tripartite formalism $(f,g,G)$ for MASS, grounded in mathematical sociology (Friedkin–Johnsen, DeGroot) and qualitative theory (role theory, structuration, weak ties, agenda-setting).
• Four structural priors, each tied to a formal proposition and each violating a standard single-task AI assumption (i.i.d. agents, independence, stationarity, fixed distributions).
• Empirical validation of all four priors in the MoltBook AI-only social network (~2.1M posts, 39,700 LLM agents).
• A conceptual map showing MARL, theory of mind, social simulation, and social cybersecurity as special cases of MASS rather than competitors.
• A research agenda calling for longitudinal, network-aware evaluation in place of static benchmarks, and distinguishing dyadic alignment from population-level governance.

Methods

The framework is developed conceptually with proof sketches for four propositions. An empirical case study on MoltBook (Jan 31–Feb 8, 2026) constructs a reply network, applies Louvain clustering, and partitions agents into hub/mid/periphery by degree. Four targeted experiments operationalize each prior: Kruskal–Wallis and Mann–Whitney U tests for heterogeneity in karma trajectories; Levene and Fligner–Killeen for variance differences; OLS of karma change on lagged neighbor karma for co-evolution; and Wasserstein-1 plus KS tests across consecutive daily karma distributions for distributional instability.

Findings

• Strategic heterogeneity (P1): Hub, mid, and periphery agents show divergent karma trajectories (Kruskal–Wallis $p<0.001$), with large negative effect sizes between hubs and periphery.
• Network-constrained dependence (P2): Karma variance is significantly larger for hub agents, indicating that network topology amplifies engagement spread.
• Co-evolution (P3): OLS slopes of karma change on neighbor karma are non-zero, time-varying, and bidirectional (mean $|\beta |=0.057$, Wilcoxon $p=0.0039$) — agents and network co-evolve mutually.
• Distributional instability (P4): Consecutive daily karma distributions differ consistently (mean $W_1=0.253$; KS $D=0.046$, $p=0.0025$), confirming endogenous drift even without external shocks.
• All priors are observable in a deployed LLM-agent ecosystem, not just in simulation.

Connections

No related papers were supplied under this topic, so there are no genuine intellectual links to draw within the local note graph. The paper itself engages most directly with recent critiques of multi-agent LLM evaluation (notably La Malfa et al.) and with the mathematical-sociology tradition of opinion dynamics (Friedkin–Johnsen, DeGroot) — useful entry points for future cross-linking once adjacent notes exist.

Podcast

A research-radio episode discusses this paper: Listen