AIA Model: Agency, Interaction & Adaptation

• The AIA Model is a comprehensive framework that defines 'agency' as a system's capacity to autonomously generate, select, and regulate actions toward specific goals.
• It systematically quantifies interactions using experimental paradigms and information-theoretic metrics to measure emergent behaviors like rapport and synchronization.
• It prescribes design guidelines and adaptive feedback mechanisms for optimizing human-AI collaboration, evolutionary adaptation, and collective intelligence.

Agency, Interaction, and Adaptation Model (AIA Model) describes a comprehensive framework for analyzing, designing, and evaluating complex interactive systems—especially agentic AI, human-AI collaboration, evolutionary adaptation, and collective intelligence. Under this paradigm, "agency" denotes the endogenous capacity of a system (artificial or natural) to generate, select, and regulate actions with respect to goals; "interaction" formalizes the multi-directional exchanges (information, influence, feedback) among agents and environments; and "adaptation" comprises the mechanistic feedback loops that reconfigure system behavior or structure in response to outcome discrepancies or environmental variation. This model supports theoretical unification across engineered and biological domains and anchors experimental manipulation, quantitative metrics, and practical design prescriptions.

1. Foundational Concepts and Theoretical Constructs

The AIA Model is underpinned by formal definitions developed in multiple domains. A core systems-theoretic instantiation defines an agentic system as the tuple:

$$S = (H, A, E, T, \mathfrak{F}_H, \mathfrak{F}_{AE}, \mathfrak{F}_{AA})$$

where $H$ (humans), $A$ (AI agents), $E$ (environment), and $T$ (toolsets) are coupled through explicit feedback and interaction kernels $\mathfrak{F}$, embedding each agent's functional agency via: (i) action generation policy $\pi$, (ii) an outcome model $m$ for prediction, and (iii) an adaptive update rule for $\pi$ or $m$ in response to observed discrepancies (Miehling et al., 28 Feb 2025).

Barandian systems (BDI), control-theoretic architectures, and algorithmic frameworks such as the Phenopoiesis Algorithm provide alternative formalizations of agency, self-organization, and interactive adaptation (Dignum et al., 21 Nov 2025, Le, 1 Feb 2026, Eslami et al., 11 Mar 2026). Across these models, the feedback between internal generative structure, state coarse-graining, and top-down constraints is recognized as foundational to "agency" in both biological and engineered systems (Horibe et al., 7 Dec 2025).

2. Experimental Paradigms and Quantitative Metrics

Experimental realization of the AIA Model proceeds from controlled manipulations of agency and adaptation variables. In companion chatbots, a 3×2 factorial design manipulated user authorship (avatar generation) and linguistic style matching (static vs. adaptive), revealing that user-directed, visible agency (avatar control) robustly elevates rapport, while high-frequency, covert mimicry destabilizes personalization and satisfaction—termed the Adaptation Paradox (Brandt et al., 16 Sep 2025). Key statistical outcomes include $\omega^2=0.040$ for rapport (p=0.013) and negative effects of adaptive LSM on personalization and satisfaction (Cohen's $d=0.35\text{–}0.48$).

Information-theoretic metrics such as bi-predictability ($P$), defined as the mutual information between joint (state, action) and next state divided by the total interaction entropy, quantify the degree of shared structure in agent-environment dynamics. $P = 1$ for quantum-correlated systems, classically bounded by $P\leq0.5$ and further reduced ($P\approx 0.33$) by the introduction of agency (Hafez et al., 26 Feb 2026).

In agent-based organizational models, controlled agent learning $(\mathcal P)$ and group adaptation intervals $(\tau)$ allow measurement of non-linear, often underproportional, interactions between individual exploration and collective recombination, with interaction coefficients $IE<1$ signaling attenuation due to interdependencies (Blanco-Fernandez et al., 2022).

3. Mathematical Structures and Feedback Mechanisms

Core mathematical representations include multi-agent POMDPs interconnected through message-passing and tool-invocation (Miehling et al., 28 Feb 2025). Feedback adaptation is operationalized at multiple nested levels:

• Environment–Agent (predictive-coding): State updates $$s_{t+1} = s_t + \eta_s \sum_\mu W^\mu \varepsilon_i^{t,\mu}$$ align internal states with multi-modal prediction errors.
• Agent–Environment (free energy minimization): Each agent minimizes a free-energy functional, evolving causal models and supporting active inference for interventional reasoning.
• Agent–Agent (collective metacognition): Consensus confidence $\mu_G^t$ and group uncertainty drive social Bayesian updating of parameters.

Control-theoretic generalizations embed agency within layered feedback loops, where increasing agentic authority introduces mechanisms such as parameter adaptation, endogenous mode switching, decision-induced delays, and on-the-fly reconfiguration of both controller objectives and architectures (Eslami et al., 11 Mar 2026).

4. Case Studies: Embodiments and Empirical Outcomes

Empirical work substantiates the model in several contexts:

• Embodied Conversational and Musical Agents: Finite-state turn-taking dynamics, modulated by interpersonal dominance ($D_{ij}$) and liking ($L_{ij}$), yield emergent synchrony, quantifiable via phase-locking indices and mutual information measures. In pilot studies, phase synchrony $S_{phase}$ improved to $0.73 \pm 0.05$ and subjective ratings increased with adaptive turn-taking (Sanlaville et al., 2015).
• Evolutionary Adaptation (Phenopoiesis): Organisms with dual (genetic and epigenomic) inheritance recover from environmental shifts 3.4 times faster than gene-centric baselines and sustain multi-task performance ($91.2\%$, $p<0.0001$ vs. control). Online decision-theoretic control of exploration–exploitation and phenotype-to-epigenome write-back channels realize active, compositional agency (Le, 1 Feb 2026).
• Human–AI Collaboration: The Agency–Interaction–Adaptation conceptual model was iteratively validated through literature review and expert interviews, supporting a 9-dimensional design space with discrete metrics for agency distribution, allocation, interaction intent, guidance, and adaptation method (Holter et al., 2024).
• Human-Robot Interaction: Systematic review identifies adaptiveness, communication style, anthropomorphism, presence, and user differences as critical factors influencing perceived autonomy and agency, as measured by psychometric scales and intentional binding paradigms (Glawe et al., 26 Sep 2025).

5. Theoretical Extensions and Spectrum of Agency

Recent theoretical work distinguishes between structural (designer-imposed) and teleological (self-generated constraint) agency. The Agency Index aggregates the magnitude of the internal predictive gap (internal regulation) and the irreducibility of intrinsic dynamics to an external observer’s model (external divergence), providing a normalized measure ($\mathcal{A}\in[0,1]$) of where a system lies on the agency continuum (Horibe et al., 7 Dec 2025).

Control-theoretic analysis formalizes agency as hierarchical decision authority. The agency ladder spans from fixed, reactive rules (Level 0) to endogenous objective generation and architectural reconfiguration (Level 4), with each stage introducing escalating dynamical complexity—time-varying adaptation, switched/hybrid systems, and stability–safety challenges addressable by Lyapunov and quadratic-cost arguments (Eslami et al., 11 Mar 2026).

6. Design Implications and Prescriptive Guidelines

Design guidelines derived from empirical and theoretical work emphasize:

• Prioritizing user-driven, visible personalization (e.g., co-creation of avatars or persona settings).
• Stabilizing system-driven adaptation: smoothing style shifts, filtering out ephemeral or incoherent mimicry.
• Enhancing legibility: surfacing adaptation cues (e.g., “adjusting tone…”), and affording user control over adaptive behaviors (Brandt et al., 16 Sep 2025, Yun et al., 30 Jan 2026).
• Instrumenting feedback loops with real-time agency/affordance tracking via neurodynamic and phenomenological markers in XR/BCI, using weights calibrated to task demands (Hila, 9 Sep 2025).
• Structuring human–AI collaborations with explicit metrics for initiative, negotiation, and role allocation, supporting quantitative analysis and targeted design interventions (Holter et al., 2024).
• For evolutionary domains, enabling fast phenotypic learning to couple with slower genetic change and permitting cross-generational compositional reuse (Le, 1 Feb 2026).

7. Open Challenges and Research Directions

Key outstanding questions addressed in the literature include:

• Scaling generalist, curiosity-driven agency while balancing efficient task decomposition and delegation (Miehling et al., 28 Feb 2025).
• Monitoring, controlling, and governing emergent subgoal formation and distributed authority.
• Operationalizing and standardizing measurement of autonomy, agency, and adaptation—via both phenomenological and information-theoretic constructs (Glawe et al., 26 Sep 2025, Hafez et al., 26 Feb 2026).
• Advancing from systems with agency (choice, effect, asymmetry) to systems with true intelligence (learned causal models, self-monitoring, adaptive scope) by closing the loop with real-time feedback architectures (Hafez et al., 26 Feb 2026).
• Disentangling structural and teleological dynamics in both single and multi-agent settings, foregrounding generative divergence and emergent collective agency (Horibe et al., 7 Dec 2025).

---

Across cognitive science, organizational theory, AI, control, and evolutionary biology, the Agency–Interaction–Adaptation Model unifies the analysis of action, feedback, and learning. It provides both a taxonomic framework and operational scaffold for understanding and engineering systems that are adaptive, interactive, and possessed of genuine or functional agency.