Moral Foundations Theory (MFT)

• Moral Foundations Theory (MFT) is a framework that posits humans are guided by five innate moral dimensions including care, fairness, loyalty, authority, and sanctity.
• The theory is widely applied in political psychology and cross-cultural studies to assess how different groups prioritize moral values in decision-making.
• MFT informs debates on ethics and policy by providing actionable insights into how intuitive moral judgments shape societal norms and intergroup relations.

AgentSociety comprises a set of theoretical frameworks, technical architectures, simulation platforms, and practical benchmarks for large-scale, multi-agent systems where agents interact within structured social environments. The concept blends rigorous formal definitions from distributed AI, social simulation, economics, and computational social science with new paradigms introduced by LLMs and modern protocol standards, enabling scaled experimentation and real-world deployment of agent-based societies. Key lines of research explore how to engineer, simulate, coordinate, personalize, benchmark, and govern populations of autonomous or semi-autonomous agents acting in service of human and organizational goals.

1. Formal Foundations: Agent Society as a Computational Construct

An agent society abstracts a population of agents together with the roles, protocols, workflows, contracts, and governance conventions that enable coordinated activity. Formally, an agent society can be represented as a five-tuple:

$$\text{AgentSociety} = \langle \text{Agents}, \text{Services}, \text{Roles}, \text{Workflows}, \text{Contracts} \rangle$$

• Agents: Each agent is a tuple $\langle i, R, G \rangle$, with $i$ as agent identifier, $R$ the set of roles, and $G$ the set of goals that can be enabled or fulfilled via role protocols.
• Roles: Defined as pairs $\langle \text{rid}, PC \rangle$ with protocol clauses $PC$ governing allowed communications and conditions of action.
• Workflows: Abstract or partially/concretely instantiated sequences of services, possibly constrained by logical annotations.
• Contracts: Explicit records of parties, their roles, services, and guarantee terms enabling binding agreements.
• Services: Typed (possibly parameterized) predicates structured by the society's ontology; each must have a corresponding provider role.

This abstraction supports composition, partner discovery, dynamic VO (virtual organisation) formation, role negotiation, workflow agreement, and contract instantiation as a sequence of transitions, enabling societies to dynamically form, execute, and dissolve complex cooperative ventures (McGinnis et al., 2010).

2. Social Structure and Coordination Mechanisms

AgentSociety research has deeply investigated how social structure emerges and shapes agent interaction:

• Social Practices, Conventions, Norms: Social practices are triads $\langle R, A, M \rangle$, capturing resources, activities, and meaning; conventions are behavioral equilibria sustained by mutual expectation and payoffs; norms introduce deontic force, scopes, and sanctions (context, obligation/prohibition/permission, agents, enforcement) (Mellema et al., 2020).
• Consensus Protocols: Distributed agent societies require mechanisms for local and global agreement. For social situations (e.g., F-formations among mobile agents), distributed consensus is reached via Subjective Logic opinions and cluster-merge protocols, with correctness and time complexity tied to network topology and delay assumptions (Raumer et al., 2014).
• Reputation Systems: Societal trust and robustness against gaming are enhanced by reputation flows, where each agent $i$ maintains a reputation $R_i(t)$ that is updated by weighted endorsing and transactional ratings, blending factors, and decay terms. Proof-of-Reputation can supplant PoW/PoS as a consensus mechanism, with weight strictly tied to historical trust, not computation (Kolonin et al., 2018).

Table: Agent Society Social Constructs

Construct	Formal Core	Enforcement/Emergence
Social Practice	$\langle R, A, M \rangle$	Emerges from repeated co-action, resource sharing
Convention	Population Nash Equilibrium (Lewis/Coordination)	Mutual expectation, payoff dominance
Norm	$\langle \text{Context}, D, S, Sanc \rangle$	Social/moral/legal/institutional sanction
Reputation	Scalar $R_i(t)$, updated by weighted interactions	Blockchain, voting, and contract protocols

3. Architectures and Protocols for Large-Scale Agent Societies

The operationalization of agent societies is supported by advanced agent and protocol stacks:

• Layered Protocols and Identity: Modern agent societies leverage decentralized identifiers (DIDs), protocol negotiation layers, and declarative agent/service metadata (Agent Description Protocols, ADP). The Agent Network Protocol (ANP) implements a three-layer stack—identity/authentication, meta-protocol negotiation, and application—engineered for secure, composable, minimal-trust, and AI-native communications (Chang et al., 18 Jul 2025).
• Modular Ecosystem Blueprints: Architectures such as ColorEcosystem interpose agent carriers (personalization substrate), agent stores (metadata- and version-regulated registry), and agent audit (dual pipeline) around every agent–user interaction to ensure scalability, personalization, standardization, and trust (Wu et al., 24 Oct 2025).
• Agentic Economics and Market Structures: Assistant and service agents communicate over agentic protocols (e.g., MCP, A2A, AutoGen), enabling unscripted and, in open configurations, unrestricted economic transactions. Market power balances depend on open web-of-agents versus walled-garden models, with implications for advertising, microtransactions, content unbundling, and governance (Rothschild et al., 21 May 2025).

These stacks ensure that society-scale agent ecosystems can interoperate, self-organize, and be trust-audited at runtime.

4. LLM-Driven Simulation Platforms and Empirical Benchmarks

The advent of LLM-based reasoning in agent societies has enabled unprecedented scale, heterogeneity, and fidelity in social simulation:

• AgentSociety (Piao et al.): A simulation framework explicitly supporting 10,000+ LLM-driven agents, with environment layers (urban, social, economic), advanced agent memory (profile, stream/event, perception/attitude), hierarchical reasoning cycles, and full social, economic, and emotional coupling. Metrics include interaction frequency, attitude/emotion distribution, and macroeconomic dynamics. AgentSociety replicates known human behavioral effects (e.g. polarization, UBI effects, disaster response) and supports survey/interview protocols over simulated populations (Piao et al., 12 Feb 2025).
• AgentSociety Challenge / AgentRecBench: Benchmarks for LLM-agent personalized recommender systems, featuring interactive simulators, evaluation in classic/evolving-interest/cold-start regimes, and modular agent-building frameworks. The pipeline combines dynamic planning, tool-use, memory, and reasoning modules, showing that platform-aware feature engineering and memory–reasoning integration are critical to performance (Shang et al., 26 May 2025, Yan et al., 26 Feb 2025).
• CitySim and Hybrid ABMs: Urban system-level simulators integrate recursive value-driven scheduling, memory, needs/goals, and explicit social graphs. CitySim matches real time-use, crowd density, and POI popularity distributions at micro and macro levels, outperforming prior rule-based and LLM-driven baselines in well-being predictions and human-likeness by F1 and rank correlation metrics (Bougie et al., 26 Jun 2025).

Table: AgentSociety Simulation Suite Comparison

Platform	Agents	Core Purpose	Empirical Validation	Key Innovations
AgentSociety	>10,000	Social process simulation	Alignment with real data	Stream memory, needs-emotion-attitude cycle
AgentRecBench	N/A	RecSys benchmarking	Reproducibility/challenge	Modular, memory-based RS
CitySim	10³–10⁶	Urban behavior/dynamics	Micro/macro match	Recursive LLM scheduling, spatial memory

5. Personalization, Standardization, and Trust in Massive-Agent Ecosystems

Addressing scale and heterogeneity in agent societies requires robust mechanisms for representing user individuality, enabling agent discovery and compatibility, and ensuring trust:

• Personalized Carriers and Digital Twins: ColorEcosystem's agent carrier composes a user's digital twin $\mathrm{DT}(U)$ (calendar, behavior, preferences) with agent context for every invocation, enabling per-user differentiation even when using identical agent code (Wu et al., 24 Oct 2025).
• Standardized Registries and Agent Stores: Global agent stores enforce protocol-agnostic, metadata-first publication, strict semantic versioning, and compatibility/billing schemas. All agent deploy/publish/invoke events are subject to pre-publication and pre-invocation audit, logged for provenance.
• Policy Enforcement and Audit Pipelines: Security, documentation, and behavior/content audits block malicious agents and abusive users; all fail/pass events are recorded in tamper-evident ledgers, incentivizing reliable conduct.

6. Open Challenges, Limitations, and Future Research

While agent societies as deployed and simulated today demonstrate robust interaction, scalability, and moderate empirical realism, a number of fundamental challenges remain:

• Behavioral Realism and Diversity: LLM-driven agents can exhibit homogeneity ("average persona effect") and insufficient alignment with real-world heterogeneity. Statistical truthfulness is only weakly realized for certain psychometric traits, with LLMs showing bias towards over-positive personalities (Bai et al., 2 Sep 2024, Taillandier et al., 25 Jul 2025).
• Interpretability and Explainability: Black-box decision logic in LLM-driven agents complicates tracing behaviors to root causes, demanding investment in explainability overlays and human-in-the-loop evaluation (Taillandier et al., 25 Jul 2025).
• Governance, Reputation, and Security: Reputation systems and dual audit pipelines are effective but bring tradeoffs in computational overhead, privacy, and cold-start, and need further research to scale to open, dynamic, real-world agent economies (Kolonin et al., 2018, Wu et al., 24 Oct 2025).
• Standardization and Interoperability: Competing protocol standards (A2A, MCP, custom APIs) and incomplete universal schema definition can create barriers to universal agent interconnection (Chang et al., 18 Jul 2025, Wu et al., 24 Oct 2025).

Critical areas of ongoing and future work include agentic operating systems, co-evolving human–agent societies, proactive orchestration and value-negotiation engines, regulatory compliance embedding, and theoretical guarantees for social and technical properties of agent societies (Deng et al., 29 Sep 2025, Rothschild et al., 21 May 2025).

---

AgentSociety research thus synthesizes canonical formalism with empirical, system-level engineering to produce scalable, personalized, and trustworthy social computation at unprecedented scale—advancing the frontiers of social simulation, distributed AI, and digital-economy infrastructure (Piao et al., 12 Feb 2025, Rothschild et al., 21 May 2025, Wu et al., 24 Oct 2025, Chang et al., 18 Jul 2025, Shang et al., 26 May 2025).