Collective Agency: Emergent Dynamics

• Collective Agency is the emergent capacity of groups to act as unified entities with properties beyond the sum of individual actions.
• Formal frameworks apply logical, game-theoretic, and information-theoretic models—using metrics like Pareto optimality, Nash equilibrium, and Markov blankets—to quantify CA.
• Applications of CA span AI alignment, multi-agent adaptive systems, and biological collectives, demonstrating its broad practical and theoretical significance.

Collective agency (CA) denotes the emergent capacity of a group, system, or ensemble to act as a unified agent, exhibiting properties, decision-making, or outcomes distinct from the mere aggregation of its individual components. In contemporary research, CA is formalized variously: as logical coordination among agents, information-theoretic emergence of macro-level autonomy, or as an alignment principle for artificial intelligences transcending fixed individualist objectives. Recent work spans logics of dependence, distributed adaptive intelligence, group agency in multi-agent inference, logics of group cohesiveness, and dynamic alignment frameworks for LLMs.

1. Theoretical Foundations of Collective Agency

Multiple theoretical traditions converge on the concept of collective agency:

• Logical and Game-Theoretic Accounts: CA is formalized through preference- and dependence-logics that abstract group capacities independent of individual intentionality. Notably, the Logic of Preference and Functional Dependence (LPFD) recasts collective action in terms of Pareto optimality and Nash equilibrium, using syntactic operators for functional and preference interdependence (Shi et al., 2021).
• Performance and HCI Perspectives: In interactive systems, CA is interpreted as an emergent, relational phenomenon. The DANCE² project frames CA not as simple action aggregation but as a co-constructed property of participants, technology, and choreographic design, emphasizing the negotiation among agentive behavior (observable action), felt agency (subjective sense), and actual power (Sathya et al., 11 Jun 2025).
• Active Inference and Information-Theoretic Emergence: The emergence of CA is also modeled in multi-agent active inference systems, e.g., bird flocks, where mutual coupling yields a higher-order Markov blanket encapsulating a new agent (the flock), capable of coordinated inference and action beyond the capacity of any member (Maisto et al., 13 Nov 2025).
• Embodied Adaptive Intelligence: In the context of artificial intelligence, CA underpins “collective adaptive intelligence”—a structural, dynamical reconfiguration of agent ensembles yielding robust, scalable, and task-general behavior (Wang et al., 29 May 2025).
• Logical Accounts of Group Cohesiveness: New logics model group agency by encoding pro-social relations (“assistance”) among coalitions—defining when a group can cohesively “bring about” a goal given a class of cohesion networks (Troquard, 2 Nov 2025).

2. Formal Frameworks and Metrics

Formalization of collective agency varies by paradigm:

Logical and Game-Theoretic Models

LPFD, as detailed in (Shi et al., 2021), introduces a syntax:

• $$\llbracket =_{X}, \preceq_Y, \prec_Z \rrbracket \varphi$$: evaluates properties over action profiles, capturing interdependence and Pareto efficiency.
• Pareto optimality and Nash equilibrium have explicit semantic and syntactic characterizations enabling precise reasoning about when a coalition exhibits CA.

The core definition of CA for a group $X$ requires:

• Joint strategy profile is a Nash equilibrium ($NE_X$).
• Every subgroup in some cover $\mathcal{C}$ of $X$ is internally Pareto-optimal when actions of $-X$ are fixed.

Cohesion Networks and Logics of Assistance

(Troquard, 2 Nov 2025) defines a cohesion network as a graph whose vertices are strict subgroups and edges prescribe pro-social (“successful assistance”) behaviors. The extended BIAT language introduces:

• $[C_1 \rightarrow C_2]\varphi$: agent(s) in subgroup $C_1$ ensure that $C_2$, in trying for $\varphi$, succeed.
• $E_G\varphi$: group $G$ brings about $\varphi$ iff all proscribed assistance edges are realized. This reductionist but expressive framework yields a family of decidable logics parameterized by network class (e.g., acyclic vs. strongly connected cohesion).

Active Inference and Information Theory

(Maisto et al., 13 Nov 2025) shows that sufficiently coupled agents form emergent Markov blankets, enabling the definition of a macro-agent (the flock) whose internal, sensory, and active states can be exactly specified. The transition from micro- to macro-level agency is denoted by:

• Emergent Markov boundary partitioning.
• Macro-level free energy minimization dynamics.
• Synergistic information (via PID): the flock encodes information (about, e.g., predator position) inaccessible to any constituent agent—a direct quantification of implicit collective knowledge.

AI Alignment and Dynamic Objective Functions

For alignment of advanced AI systems, CA is operationalized as an alignment direction, not a scalar metric. In (Anantaprayoon et al., 5 Dec 2025), CA is instantiated as:

• “Infinite expansion of agency across spacetime.”
• Operationally via a self-reward function: responses are scored on a 0–5 scale reflecting balanced advancement across knowledge, power, vitality, and benevolence.
• Policy optimization uses Group Relative Policy Optimization (GRPO) in a self-improving loop without external human feedback.

3. Design Principles, Attributes, and Empirical Observations

Distinct research traditions attribute varying design imperatives and empirical metrics to CA:

System Properties in Embodied and Adaptive AI

(Wang et al., 29 May 2025) identifies four key attributes of CA in adaptive collectives:

• Task/Topology Adaptation: generalization to OOD tasks and novel communication graphs.
• Resilience: graceful degradation under up to $k$ agent failures.
• Scalability: collective performance grows with ensemble size on complex tasks.
• Self-Assembly: convergence to function-suited communication topologies.

A canonical agent-update rule incorporates self-adaptation, role negotiation, and decentralized communication constraints.

Participation, Agency, and Power in Interactive Systems

(Sathya et al., 11 Jun 2025) empirically disentangles agentive behavior (actions taken), felt experience of agency (subjective sense), and actual power (outcome impact) in audience–robot–dancer systems. Metrics include:

• Override ratio $R_{p,i}$ at each decision point.
• Clustering of collective decision patterns (measured by coefficient of variation).
• Participants’ self-reported feelings of agency and connection.

Information-Theoretic Manifestation in Biological Collectives

Empirical emergent effects in agent collectives are captured via:

• System-wide response latency to perturbations.
• Energy functions indicating global alignment.
• Synergy (PID) quantifying information present only at the collective level (Maisto et al., 13 Nov 2025).

4. Philosophy, Conceptualization, and Extensions

Historical philosophical models typically invoke “we-intentions” (Bratman, Searle, Tuomela). Modern formalizations replace intentional primitives by logical, informational, and structural constructs. For example, LPFD’s quantification over covers and subcoalitions encompasses nuanced forms of agency, moving beyond individual best-response logic (Shi et al., 2021).

Current frameworks highlight the distinction between:

• Mechanistic agency: emergent from constraint and structural coupling.
• Normative/ethical dimensions: Who is empowered? Is agency orchestrated or co-created?
• Temporal/dynamic extensions: Most formal logics to date are static, though causality and temporal commitment are areas of active investigation.

A plausible implication is that future frameworks will integrate epistemic, deontic, and dynamic modalities to model the full spectrum of collective intentionality, learning, memory, and normativity.

5. Applications Across Domains

CA underpins applications in AI, HCI, sociotechnical systems, and biological modeling:

• AI Alignment: CA serves as a scalable, open-ended alignment direction for advanced models, operationalized via self-improving mechanisms harnessing model-generated feedback and multi-dimensional agency criteria (Anantaprayoon et al., 5 Dec 2025).
• Multi-Agent Embodied Intelligence: CA architectures allow adaptive, robust performance in dynamic settings with changing system size/topology (Wang et al., 29 May 2025).
• Performance and Interactive Art: CA can be staged and observed in choreographed human-robot-audience configurations, informing participatory system design (Sathya et al., 11 Jun 2025).
• Biological Systems: CA models explain the emergence of macro-scale adaptive behaviors and implicit collective knowledge in flocks, neural assemblies, or tissues (Maisto et al., 13 Nov 2025).
• Group Social Dynamics: Cohesion networks and BIAT-based logics provide analytic tools for modeling collaboration, assistance, and social fabric in group tasks (Troquard, 2 Nov 2025).

6. Limitations, Open Problems, and Future Directions

Despite substantial advancements, current approaches to CA have notable limitations:

• Most logical frameworks are temporally static, missing process dynamics and longitudinal role changes (Shi et al., 2021, Troquard, 2 Nov 2025).
• Many models, especially in conceptual adaptive intelligence, present blueprints without full empirical realization or benchmarks (Wang et al., 29 May 2025).
• Open-endedness of CA as an alignment principle is incompletely formalized; the interplay among its entangled aspects invites further interpretability and metric development (Anantaprayoon et al., 5 Dec 2025).
• Group agency models in performance and HCI settings illuminate gaps between experience and effect, but real-world algorithmic mediation under complex feedback remains challenging (Sathya et al., 11 Jun 2025).

A plausible implication is that future CA research will deepen integration of temporal, normative, and epistemic modalities, embrace richer statistical and information-theoretic metrics, and more rigorously link empirical benchmarks to formal structure. Multi-agent negotiation, decomposition of agency energies, and curriculum-driven emergent capabilities are expected to drive progress in both theoretical and applied contexts.