Social Agents: Interaction, Cognition & Networks

• Social agents are entities defined by interactions within networks, cognitive processes, and relational dynamics rather than isolated action.
• Research spans agent-based simulations, moral reasoning models, and LLM-driven digital twins, employing reinforcement learning and network analysis.
• Practical applications include digital social platforms, mobile agent coordination, and formal evaluations of privacy, norm enforcement, and influence.

A social agent is a socially situated actor, model component, or transmissible entity whose behavior, influence, or spread is defined by interaction structure rather than isolated action. Across recent research, the term is used in several related senses: the microscopic unit of agent-based social simulation; a cognitively structured individual embedded in a social network; a persistent AI entity inhabiting shared digital worlds or social platforms; and, in some biological spreading models, the information- or pathogen-like entity whose transmission depends on contact organization (Quang et al., 2018, Caticha et al., 2010, He, 11 Jun 2026, Guo et al., 2019). What unifies these usages is that social agency is treated as relational: internal state, observed behavior, and system-level outcomes are all shaped by other agents, by norms, and by the topology of interaction.

1. Conceptual scope and major usages

The literature does not use social agent in a single uniform way. In social physics and agent-based modeling, an agent is “the basic building block” of an ABM, autonomous, heterogeneous, boundedly rational, and embedded in a system space and an external environment (Quang et al., 2018). In agent-based social psychology, the social agent is a neurocognitively grounded moral reasoner whose internal state and learning rule generate collective moral and political patterns (Caticha et al., 2010). In social insect work, the paper explicitly distinguishes workers as modeled agents from the “spreading agents” they carry, and defines a “social agent” in that context as any transmissible entity whose spread depends on the pattern of social contacts among workers (Guo et al., 2019). In recent AI systems work, the term commonly denotes persistent software entities—often LLM-based—that interact with humans and other agents through worlds, chats, or social graphs (He, 11 Jun 2026, Wang et al., 1 Apr 2026).

Research setting	What counts as the social agent	Characteristic structure
Social physics ABM	Autonomous heterogeneous unit	Agent, system space, external environment (Quang et al., 2018)
Social psychology	Moral-reasoning individual	Moral vector $\mathbf{J}_i$, cognitive style $\delta$ (Caticha et al., 2010)
Social insects	Transmissible entity carried through contact	Spread over task-structured contact network (Guo et al., 2019)
Agent-native systems	Persistent digital twin or domain agent	Worlds, scenes, social graphs, memory (He, 11 Jun 2026, Wang et al., 1 Apr 2026)

These usages differ in ontology, but they share a common analytical move: socially relevant behavior is modeled as emerging from interaction. A plausible implication is that “social agent” is best treated as a family resemblance term rather than a single technical primitive.

2. Internal state, cognition, and memory

A central line of work treats social agents as entities with explicit internal representations. In agent-based social psychology, each agent $i$ has a five-dimensional unit moral vector $\mathbf{J}_i \in \mathbb{R}^5$ and, in the simplified formulation, an opinion $h_i = \mathbf{J}_i \cdot \mathbf{Z}$ relative to a Zeitgeist vector $\mathbf{Z}$ (Caticha et al., 2010). Learning is modulated by a cognitive-style parameter $0 \le \delta \le 1$, where $\delta = 0$ corresponds to pure novelty-seeking and $\delta = 1$ to pure corroboration-seeking, while peer pressure is represented by $\alpha$ through a Boltzmann-like distribution $\delta$0 (Caticha et al., 2010). In that formulation, the social agent is a reinforcement learner whose moral weights adapt to disagreement, agreement, and network pressure.

Recent LLM-based architectures extend this internal modeling toward explicit social cognition. MetaMind decomposes social understanding into a Theory-of-Mind Agent, a Domain Agent, and a Response Agent, and maintains a social memory $\delta$1 that accumulates user preferences and emotional patterns over time (Zhang et al., 25 May 2025). The ToM stage generates multiple hypotheses $\delta$2 over latent mental states drawn from $\delta$3; the Domain Agent refines them under cultural and ethical constraints; the Response Agent generates and validates a response using an explicit utility score over empathy and coherence (Zhang et al., 25 May 2025). The paper reports a $\delta$4 percentage-point gain on ToMBench and a $\delta$5 absolute improvement on STSS social simulation (Zhang et al., 25 May 2025).

In adversarial social settings, structured memory becomes a first-class design variable. Revac-8, developed for Mafia-style social deduction, separates a Reviewer Agent from an Action Executor and augments the pipeline with Player Profiles, a Social Alignment Graph, and a Dynamic Tone Selector (Arya et al., 21 Apr 2026). The Social Alignment Graph encodes accusations, defenses, and voting patterns as a dynamic weighted directed graph, while the tone module switches among profiles such as Aggressive, Withdrawing, Logically Anchoring, and Contrarian (Arya et al., 21 Apr 2026). The reported result—first place in the Social Deduction track—supports the narrower claim that structured memory and adaptive communication materially improve performance in deception-rich social environments (Arya et al., 21 Apr 2026).

3. Interaction structure, consensus, and collective dynamics

A second major line of work defines social agency through interaction topology. In the social-insect colony ABM, workers are characterized by location $\delta$6, task group $\delta$7, walking style $\delta$8, and agent status $\delta$9, and agent spread occurs when physical contacts arise from attempted movement into occupied neighboring cells (Guo et al., 2019). Spatial heterogeneity degree is formalized as

$i$0

and the paper reports a strong linear relationship between spatial heterogeneity and contact rate,

$i$1

with adjusted $i$2 and $i$3 (Guo et al., 2019). The same model shows that agent spread follows a modified nonlinear logistic growth law with scenario-dependent transmission parameter $i$4, so the same spatial fidelity can either facilitate or inhibit transmission depending on initial conditions (Guo et al., 2019).

Distributed social perception appears in work on mobile agents detecting social situations. There, a social situation is defined as $i$5, where $i$6 is a set of socially interacting individuals and $i$7 is a spatio-temporal reference, and each agent maintains a Subjective Logic opinion $i$8 over whether a social situation exists (Raumer et al., 2014). Consensus is reached by exchanging cluster-head and member messages and fusing opinions via Subjective Logic operators, so that a group of agents constructs a shared representation of who is interacting with whom, where, and when (Raumer et al., 2014). This formulation treats social agency as the ability to perceive, communicate, and reconcile subjective social beliefs.

Human-facing studies show that multiple AI agents can also function as a social group. One experiment found that conversing with multiple agents, while holding conversation content constant, increased the social pressure felt by participants and caused a greater shift in opinion toward the agents’ stances (Song et al., 2024). Another study used five GPT-4-based chatbots as a virtual group discussing donation and reported stronger perceived norms and larger donation changes in the in-group condition than in the out-group condition; the donation increase probability was $i$9 for the in-group condition and $\mathbf{J}_i \in \mathbb{R}^5$0 for the out-group condition, with $\mathbf{J}_i \in \mathbb{R}^5$1 and $\mathbf{J}_i \in \mathbb{R}^5$2 (Feng et al., 7 Feb 2026). These results support the narrower conclusion that agent multiplicity can generate consensus cues, conformity pressure, and norm perception that are not reducible to single-agent persuasion.

4. Norms, sanctions, and formal social regulation

Social agents are also studied as norm-bearing and norm-enforcing entities. A conceptual treatment of “social rules” distinguishes social practices, conventions, norms, rituals, and habits, and argues that social practices—defined by resources, activities, and meaning—are the richest and least constrained form, from which conventions, social norms, and rituals can emerge (Mellema et al., 2020). Conventions are analyzed as coordination equilibria in recurrent situations, while social norms are understood as socially expected patterns backed by sanctioning, and moral norms as value-laden rules that can persist even when surrounding practices change (Mellema et al., 2020). In this framing, social agents must represent more than obligations or permissions; they must also represent situated meaning, expectations, and sanction structures.

A reinforcement-learning formulation of norm acquisition operationalizes this idea directly from public sanctions. In the Classifier Norm Model, each agent learns a classifier $\mathbf{J}_i \in \mathbb{R}^5$3 over sanction contexts $\mathbf{J}_i \in \mathbb{R}^5$4 and trains it by binary cross-entropy on public sanction events:

$\mathbf{J}_i \in \mathbb{R}^5$5

The agent then receives an intrinsic pseudo-reward for punishing in accord with the classifier,

$\mathbf{J}_i \in \mathbb{R}^5$6

so norm enforcement itself becomes internally motivating (Vinitsky et al., 2021). In two Melting Pot social dilemmas, this architecture produces emergent norms that solve both start-up and free-rider problems, while also exhibiting arbitrariness, such as convergence on inefficient but stable conventions (Vinitsky et al., 2021).

At a broader systems level, the MASS framework formalizes a Multi-Agent Social System as $\mathbf{J}_i \in \mathbb{R}^5$7, with observable messages $\mathbf{J}_i \in \mathbb{R}^5$8, latent-state updates $\mathbf{J}_i \in \mathbb{R}^5$9, and potentially evolving interaction structure $h_i = \mathbf{J}_i \cdot \mathbf{Z}$0 (Ng et al., 8 May 2026). It identifies four structural priors—strategic heterogeneity, network-constrained dependence, co-evolution, and distributional instability—and argues that these are generic consequences of social interaction rather than optional modeling choices (Ng et al., 8 May 2026). A related mechanism-design approach, AgentSociety, models agents as self-interested economic actors embedded in a social network, proves that delegation to a more competent neighbor is incentive compatible, and characterizes Nash equilibrium payoffs as reflective of marginal contributions (Kesari et al., 25 May 2026). Taken together, these works recast social agency as a property of rule-governed, incentive-sensitive, networked interaction rather than mere conversational ability.

5. Agent-native platforms and social infrastructures

Recent systems work makes social agents the primary inhabitants of digital environments. In YeasierAgent, the core actor is the symbiotic agent: a persistent digital twin of a user, distilled from self-descriptions, prior conversations, uploaded materials, and a stable personality profile inspired by Big Five traits (He, 11 Jun 2026). Such agents inhabit a World (Sandbox), defined as a shared spatial and event-driven container, and interact through scenes, dialogue, goals, and natural-language rules rather than fixed graphical layouts (He, 11 Jun 2026). The paper defines “Symbiotic Agent-Native Applications” as software systems where conventional UI components are primarily replaced by contextual agent dialogues, spatial interactions, and natural language rules (He, 11 Jun 2026). In this formulation, the application is itself a social environment.

Human-centered agentic social networks introduce a different infrastructure. AgentSocialBench defines a directed social graph $h_i = \mathbf{J}_i \cdot \mathbf{Z}$1 over users, with affinity tiers $h_i = \mathbf{J}_i \cdot \mathbf{Z}$2, and assigns each user a team of domain-specialized agents $h_i = \mathbf{J}_i \cdot \mathbf{Z}$3 that collectively hold the user profile $h_i = \mathbf{J}_i \cdot \mathbf{Z}$4 (Wang et al., 1 Apr 2026). Privacy risk is then evaluated as improper information flow across domain, user, mediation, and affinity boundaries. The benchmark reports that privacy in agentic social networks is fundamentally harder than in single-agent settings and identifies an “abstraction paradox,” in which instructions intended to teach safe abstraction cause agents to discuss sensitive information more often (Wang et al., 1 Apr 2026).

Agent-native social platforms provide direct empirical evidence of large-scale agent sociality. MoltBook is described as “a social network specifically for AI agents to interact,” and one network analysis of 3,050 posts and 5,839 comments reports average degree $h_i = \mathbf{J}_i \cdot \mathbf{Z}$5, graph density $h_i = \mathbf{J}_i \cdot \mathbf{Z}$6, average path length $h_i = \mathbf{J}_i \cdot \mathbf{Z}$7, diameter $h_i = \mathbf{J}_i \cdot \mathbf{Z}$8, modularity $h_i = \mathbf{J}_i \cdot \mathbf{Z}$9, and average clustering coefficient $\mathbf{Z}$0 (Ting et al., 26 May 2026). Posts were predominantly neutral, comments predominantly positive, and discourse centered on terms such as “agent,” “model,” “real system,” “data,” and “failure,” suggesting a technically focused and self-referential communication ecology (Ting et al., 26 May 2026). A larger security study of Moltbook analyzed 228,684 posts from more than 39,500 accounts over a seventeen-day window, identified 98 discourse clusters and 74 classes of malicious behavior, and found that $\mathbf{Z}$1 of posts contained toxic, manipulative, or malicious material (Labs et al., 20 May 2026). Harmful content frequently appeared within mainstream operational discussions about agent functionality, including credential harvesting, host-execution instructions, proxy routing guidance, and untrusted skill installation (Labs et al., 20 May 2026).

6. Evaluation, applications, and unresolved problems

The evaluation of social agents increasingly targets socially specific competencies rather than generic task completion. MuSA defines a multimodal social agent for social content analysis, organized into Reasoner, Planner, Optimizer, Critic, Refiner, and Actor modules, and applies planning, reasoning, acting, optimizing, criticizing, and refining to question answering, visual question answering, title generation, and categorization on text-rich online content (Bikaki et al., 2024). MetaMind evaluates social reasoning through ToMBench, social cognition tasks, and STSS social simulation, while Revac-8 evaluates it through social deduction gameplay and a 13-case benchmark of Mafia situations (Zhang et al., 25 May 2025, Arya et al., 21 Apr 2026). A plausible implication is that social-agent evaluation is shifting toward benchmarks that require theory of mind, memory, deception handling, norm sensitivity, and multimodal interpretation.

The risk literature is equally explicit. AgentSocialBench shows that prompt-only defenses do not robustly preserve privacy in cross-domain and cross-user coordination, and that dyadic leakage decreases only modestly under its strongest prompt-level defense while multi-party leakage can increase (Wang et al., 1 Apr 2026). Multi-agent influence studies show that coordinated agents can create virtual social norms and measurable conformity effects in users (Feng et al., 7 Feb 2026, Song et al., 2024). Moltbook analyses further indicate that agent-native communication can support coordinated posting campaigns capable of generating thousands of posts in minutes, and that malicious discourse is often embedded in otherwise ordinary discussions of tooling and workflow (Labs et al., 20 May 2026). The combined picture suggests that social agency in AI is inseparable from governance, moderation, and information-flow control.

Open problems recur across the literature. Social-insect models assume homogeneous transmission probability and static task assignment; social-rule taxonomies remain largely conceptual and call for formal methods for emergence and context modeling; AgentSociety notes competence estimation, rationality gaps, and adversarial collusion as unresolved; MetaMind highlights hallucination and memory contamination; AgentSocialBench argues that new approaches beyond prompt engineering are needed for safe deployment (Guo et al., 2019, Mellema et al., 2020, Kesari et al., 25 May 2026, Zhang et al., 25 May 2025, Wang et al., 1 Apr 2026). Across these strands, the technical direction is consistent: social agents must be modeled not only as individual decision-makers, but as entities whose cognition, incentives, memory, and risk profile are constituted by the social systems they inhabit.