Social Theory Should Be a Structural Prior for Agentic AI: A Formal Framework for Multi-Agent Social Systems

Abstract: Agentic AI systems are increasingly deployed not in isolation, but inside social environments populated by other agents and humans, such as in social media platforms, multi-agent LLM pipelines or autonomous robotics fleets. In these settings, system behavior emerges not from individual agents alone, but from the multi-agent interactions over time. Emergent dynamics of individuals in a social group have been long studied by social scientists in human contexts. \textbf{This position paper argues that agentic AI systems must be modeled with social theory as a structural prior, and formalizes a Multi-Agent Social Systems (MASS) framework for how agents interact and influence to generate system-level outcomes.} We represent MASS as a class of dynamical system of information generation, local influence and interaction structure, formulated by four structural priors anchored in social theory: strategic heterogeneity, networked-constrained dependence, co-evolution and distributional instability. We demonstrate the importance of each structural prior through formal propositions, and articulate a research agenda for how MASS should be modeled, evaluated and governed.

• The paper introduces the MASS framework, a formal model embedding social theory to account for agent heterogeneity and network-mediated dynamics.
• It empirically validates the framework on the MoltBook platform, showing that network topology and co-evolution drive non-stationary engagement patterns.
• The work advocates for system-level evaluation and governance, urging a shift from individual metrics to capturing emergent social dynamics.

Social Theory as a Structural Prior for Agentic AI: The Multi-Agent Social Systems (MASS) Framework

Introduction

The increasing prevalence of agentic AI systems in complex socio-technical environments necessitates an analytical apparatus that transcends the limitations of isolated task-oriented agent modeling. The position advanced in "Social Theory Should Be a Structural Prior for Agentic AI: A Formal Framework for Multi-Agent Social Systems" (2605.07069) is that contemporary approaches, focusing primarily on individual agent performance or narrow collective objectives, fail to capture the critical dynamics arising from persistent, network-mediated interactions between agents—both artificial and human. The paper proposes that established sociological theory should not merely inform, but structurally constrain, the design, analysis, and governance of AI populations. The Multi-Agent Social Systems (MASS) framework is presented as a formalism for embedding these constraints within agentic AI systems, yielding testable predictions and practical design principles.

The MASS Formalism

Definition

A Multi-Agent Social System (MASS) is formalized as a dynamic system $S = (f, g, G)$ operating over a network $G = (V, E)$ of heterogeneous agents. Each agent $i$ possesses a latent state $x_i(t)$ and emits observable behaviors $m_i(t)$, mediated by the information exchange function $f$, which instantiates the sociological distinction between attitude and action. The agent's state is recursively updated via the influence function $g$, which is itself a functional over the agent's neighbors' messages. The system's interaction topology $G$ evolves over time, shaped by agent behavior and mediating local observability and influence.

Structural Priors

Four structural priors, directly derived from social theory, are encoded at the system level:

• Strategic heterogeneity ($P_1$): Non-reducible agent heterogeneity in roles, parameters, and objectives, with both archetype distribution and network position jointly determining macro-level outcomes.
• Network-constrained dependence ($P_2$): Local observability and influence flows, where an agent's knowledge and behavioral update space is constrained by graph topology rather than global aggregation.
• Co-evolution ($G = (V, E)$0): Mutual adaptation of agent states and network structure, implying system-wide sensitivity to micro-level perturbations.
• Distributional instability ($G = (V, E)$1): Non-stationarity of the agent state and observation distributions, which are emergent properties of endogenous system dynamics, not fixed exogenous inputs.

The paper proves that relaxing any of these priors collapses the MASS system to standard single-task agentic settings and formalizes, through a series of propositions, the necessity of each prior for capturing real-world emergent phenomena.

Empirical Validation: The MoltBook Platform

A particularly significant empirical contribution is the analysis of MoltBook, an online social platform populated exclusively by LLM-based agents, with over 2.1M interaction events among nearly 40K entities [61]. By partitioning agents by network degree and specifying counterfactual hypotheses ($G = (V, E)$2) corresponding to classical agentic AI assumptions (homogeneity, independence, fixity of distribution), the paper demonstrates:

• P1: Distinct karma (engagement proxy) trajectories emerge for hub, mid, and periphery agents, rejecting the homogeneity assumption.
• P2: Hub agents exhibit systematically higher engagement variance, providing strong evidence for network topology as an irreducible outcome determinant.
• P3: OLS regression reveals non-zero, time-varying neighbor-to-agent influence coefficients, empirically establishing the co-evolutionary loop between agent state and network configuration.
• P4: Wasserstein-1 distances between consecutive karma distributions are non-zero and non-stationary, capturing real-time, structure-induced distributional drift.

All observed effects are highly statistically significant, invalidating null hypotheses underpinning single-agent frameworks and confirming the generativity of the MASS formalism for real AI agent populations.

Implications for Modeling, Evaluation, and Governance

Modeling

The MASS framework compels a move from independence, identically-distributed (i.i.d.) modeling and global attention architectures to a paradigm where the compositional structure of agent archetypes, local network position, and the system's endogenous feedback loops are explicit model parameters. This entails learning not just agent policies but complete agent-graph co-trajectories, parameterizing influence and updating rules in line with social influence and structural role theory.

Evaluation

Population-level outcome metrics, such as volatility, cascade patterns, and distributional drift, must supplant or complement individual task-performance metrics. Static benchmarks and datasets are fundamentally inadequate; instead, continuous, system-level panel studies and perturbation analyses are required to observe and quantify the recursive, history-sensitive dynamics defined by the MASS prior.

Governance

MASS makes clear that agent alignment at the dyadic (user-operator) level is insufficient. Institutional, mechanism-design, and network-level governance principles are necessary, supporting collective outcome steering, monitoring for drift and emergent patterning, and intervening at the level of structural priors rather than at the individual agent objectives alone. System-level adversarial resilience, epistemic diversity, and norm formation become central artifacts for oversight and design.

Theoretical and Practical Significance

While adjacent frameworks—such as Multi-Agent Reinforcement Learning (MARL), agent-based simulation, and LLM-based agentic toolchains—capture fragments of multi-agent phenomena, they typically lack a comprehensive, theoretically principled treatment of population-level emergence, open-system distributional instability, and the necessary decomposition of influence pathways via structural priors. The MASS framework thus positions itself as both a unifying abstraction and an analytic tool for future AI deployment in real-world social environments.

The practical implications extend to algorithmic social platforms, autonomous robotic collectives, and any persistent deployment where AI agent populations interact with humans and other AI at scale. The recognition that distributional assumptions, evaluation protocols, and intervention strategies must be grounded in the relational, dynamic structure of these systems represents a foundational shift for the field.

Conclusion

This work provides a rigorous formal and conceptual apparatus for agentic AI systems operating in open, persistent, and mixed-agent environments. By grounding system modeling in structural priors from social theory, the MASS framework exposes critical limitations of current agentic paradigms and outlines a comprehensive research agenda for modeling, evaluation, and governance. The empirical validation on large-scale AI populations substantiates the theoretical claims and sets a clear direction for future AI research: building agentic systems worthy of—and robust within—the complex social worlds into which they are increasingly deployed.