AgentSociety: Scalable LLM-Driven Agents

Updated 19 November 2025

AgentSociety is a self-organizing, scalable ecosystem of agents powered by LLMs that simulate complex social, economic, and service environments.
It employs a modular architecture that integrates agent cognition, social interaction, and hybrid decision-making to achieve empirical fidelity and scalability.
Robust protocols, decentralized identifiers, and standardized registries enable dynamic coordination and practical applications in fields like computational social science and agentic economics.

AgentSociety denotes the self-organizing, scalable ecosystem of agents—often powered by LLMs—that collectively simulate, orchestrate, and govern complex social, economic, or service environments. Distinguished from earlier agent-based systems by empirical fidelity, explicit social structures, heterogeneity of cognitive and interaction modules, and integration of standardized protocols, AgentSociety synthesizes technological, methodological, and theoretical advances across computational social science, massive-agent service platforms, and agentic economic architectures (Piao et al., 12 Feb 2025, Taillandier et al., 25 Jul 2025, Wu et al., 24 Oct 2025, Deng et al., 29 Sep 2025, Shah et al., 19 Dec 2024).

1. Architectures and Foundational Models

The architecture of AgentSociety is modular, layered, and hybrid, spanning from micro-level agent cognition to macro-level societal orchestration.

Agent Module Composition: Each agent maintains persistent memory (embedding-indexed stores or hierarchical vector DBs), structured goal sets that evolve in response to context and exogenous shocks, weighted social relations (trust, influence), and an action planner that leverages LLMs for chain-of-thought decision-making (Piao et al., 12 Feb 2025, Taillandier et al., 25 Jul 2025, Shang et al., 26 May 2025).
Environment Management: The global state consists of spatial environments (e.g., OpenStreetMap network, POIs), social graphs (mutable edge strengths), and economic entities (firms, government, banks) (Piao et al., 12 Feb 2025, Bougie et al., 26 Jun 2025).
Orchestration Layer: Discrete time scheduling, centralized or distributed, handles agent execution order and event injection. Asynchronous communication is commonplace via message brokers (MQTT, Kafka), favoring scalability (Piao et al., 12 Feb 2025, Taillandier et al., 25 Jul 2025).

Formal representation is typically $(A, E, S, T)$ , where $A$ is agent set, $E$ exogenous events, $S$ the global state, and $T: S \times A \times E \rightarrow S$ the transition function (Taillandier et al., 25 Jul 2025).

AgentSociety agents are endowed with multi-faceted cognitive modules and explicit social routines, transcending traditional reactive micro-agents.

Cognition: Each agent encompasses episodic and reflective memory, dynamic adjustment of needs and emotions (per Maslow's framework), and multi-step reasoning cycles that integrate event perception, affective updating, planning, and action execution (Piao et al., 12 Feb 2025, Bougie et al., 26 Jun 2025).
Social Interaction: Directed, weighted social graphs mediate trust, affinity, and communication frequency. Inter-agent messaging leverages LLM-informed prompts conditioned on needs, emotions, and relationship context (Piao et al., 12 Feb 2025, Shang et al., 26 May 2025).
Goal and Value Alignment: Agents adjust long-term objectives in response to unmet needs, isolation, or critical environmental changes, mimicking adaptive human reasoning patterns. Explicit querying of satisfaction indices (need-fulfillment, income-expense ratios) triggers reconfiguration (Bougie et al., 26 Jun 2025, Shang et al., 26 May 2025).
Hybridization: Critical low-level mechanics (e.g., mobility, market-clearing) are governed by ABM rules, while high-level decisions deploy LLMs for soft reasoning and negotiation (Taillandier et al., 25 Jul 2025, Deng et al., 29 Sep 2025).

3. Protocols, Interoperability, and Massive-Agent Coordination

AgentSociety necessitates robust, scalable protocols for identity, interaction, and discovery (Chang et al., 18 Jul 2025, Wu et al., 24 Oct 2025).

Identity and Registry: Agents possess decentralized identifiers (DIDs), registered via well-known endpoints (DID documents). Capabilities are published to agent directories for discovery and interoperability (Chang et al., 18 Jul 2025).
Communication and Negotiation: Agents communicate over encrypted channels, negotiate capabilities (meta-protocols) and application semantics (JSON-LD, JSON-RPC/REST, OpenAPI schemas) with adaptive protocol selection (Chang et al., 18 Jul 2025, Wu et al., 24 Oct 2025).
Massive-Agent Orchestration: Architectures such as ColorEcosystem structure agent societies into Agent Carriers (per-user, digital-twin data substrate), centralized Agent Stores (metadata-driven registries, semver versioning), and Agent Audit modules (security, behavioral, and policy enforcement for developer and user activities) (Wu et al., 24 Oct 2025).
Society Formation: Agents dynamically assemble into ephemeral or persistent coalitions, delegated service workflows, or market-based organizations, leveraging protocol-defined roles and metadata schemas for standardized, trusted interactions (McGinnis et al., 2010, Chang et al., 18 Jul 2025, Deng et al., 29 Sep 2025).

AgentSociety platforms foreground empirical validity and reproducibility, systematically benchmarking simulated outcomes against real-world data.

Scale and Performance: State-of-the-art systems simulate >10,000 agents (AgentSociety, CitySim), each sustaining memory footprints $>100$ GB and throughput $>10^6$ agent-steps/hour on cluster architectures (Piao et al., 12 Feb 2025, Bougie et al., 26 Jun 2025, Taillandier et al., 25 Jul 2025).
Validation Metrics: Calibration error, fidelity score, longitudinal consistency, and behavioral drift are computed for synthetic societies, emphasizing statistical alignment with observed phenomena (e.g., time-use surveys, mobility traces, opinion distributions) (Taillandier et al., 25 Jul 2025, Bougie et al., 26 Jun 2025).
Experimental Paradigms: AgentSociety enables controlled studies of polarization, rumor diffusion, universal basic income, and disaster response, reproducing known experimental findings and optimizing policy interventions (e.g., reach and emotional intensity of message spread, GDP and depression shifts under UBI) (Piao et al., 12 Feb 2025, Bougie et al., 26 Jun 2025).
Recommendation and User Modeling: AgentSociety Challenge and AgentRecBench demonstrate agentic recommendation benchmarks—modular LLM agents outperform classical and deep-learning methods in hit-rate and cold-start robustness (Yan et al., 26 Feb 2025, Shang et al., 26 May 2025).

Explicit modeling of social practices, roles, and norms enhances realism and regulatory compliance in AgentSociety (Mercuur et al., 2018, Mellema et al., 2020, McGinnis et al., 2010).

Social Practice Theory: SoPrA models embed hierarchical activities, context-triggered habit strengths, and shared beliefs with formal inheritance rules for value propagation (Mercuur et al., 2018).
Norms and Conventions: AgentSociety integrates deontic logic, reputation systems, and convention emergence: interaction schemas enable encoding of obligations, permissions, and sanctions; reputation-weighted consensus resist gaming and facilitate robust governance (Kolonin et al., 2018, Mellema et al., 2020).
Role/Goal/Workflow Formalization: Virtual Organizations emerge through formally defined transitions—goal identification, partner selection, role establishment, workflow agreement, and contract signing (McGinnis et al., 2010).

6. Limitations, Challenges, and Future Directions

Despite methodological advances, several limitations are acknowledged and active areas of research delineated.

Behavioral Homogeneity: LLM agents often produce “average persona” effects, suppressing minority group variance (Taillandier et al., 25 Jul 2025, Bai et al., 2 Sep 2024).
Interpretability and Drift: Black-box decision cycles challenge transparency and long-term consistency (Bougie et al., 26 Jun 2025, Taillandier et al., 25 Jul 2025).
Protocol Fragmentation and Standardization: Lack of unified registry, protocol bindings, and versioning hinders interoperability in massive-agent societies (Wu et al., 24 Oct 2025, Chang et al., 18 Jul 2025).
Value Alignment and Calibration: LLM agents skew toward social desirability, misaligning with authentic population heterogeneity—especially in psychometric distributions (Bai et al., 2 Sep 2024).

Ongoing research targets hybridizing LLMs with rule-based ABM platforms, embedding explainability and justifications, multi-modal (vision/audio) perception, scalable dynamic scheduling, multi-stakeholder value negotiation, and regulatory-aware governance (Taillandier et al., 25 Jul 2025, Deng et al., 29 Sep 2025, Wu et al., 24 Oct 2025, Bai et al., 2 Sep 2024).

7. Impact and Applications

AgentSociety underpins a broad spectrum of real-world and research applications:

Computational Social Science: Large-scale simulation platforms for reproducible, high-fidelity social experiments in opinion dynamics, economic policy, and disaster scenarios (Piao et al., 12 Feb 2025, Bougie et al., 26 Jun 2025, Taillandier et al., 25 Jul 2025).
Recommender Systems: Adaptive, agentic RSs for personalized IR, robust against cold-start and bias (Yan et al., 26 Feb 2025, Shang et al., 26 May 2025).
Agentic Economy: Democratized agent-mediated marketplaces, micro-transaction architectures, and preference-driven service ecosystems (Rothschild et al., 21 May 2025, Wu et al., 24 Oct 2025).
Service-Oriented Platforms: Self-adaptive, trustworthy, and orchestrated multi-agent systems for business process automation, collaborative problem-solving, IoT, robotics, and human-AI services (Deng et al., 29 Sep 2025, Wu et al., 24 Oct 2025).

AgentSociety thus represents the convergence of agent-based modeling, LLM-driven cognition, and protocolized service ecosystems, establishing a rigorous and extensible foundation for dynamic computational societies at scale.