Agentic AI Research Must Expand Its Paradigm to Include the Emergence of Relational Stratum

Abstract. This position paper argues that agentic AI research must expand its paradigm to include the relational stratum, comprising the intelligence that emerges in evolving human-agent systems and the dynamics observable only in their coupled operation. Contemporary AI research primarily studies intelligence at the computational stratum, comprising what models can produce, predict, solve, or optimize as artifacts in isolation. Three empirical premises, individually well-documented across the field, jointly support this position: AI agents are increasing in autonomy, they are acting on behalf of humans across longer horizons, and humans are in continuous dialogical interaction with them. Together these premises are beginning to demonstrate a co-evolutionary human-AI ecosystem with emergent properties that single-system methodology was not designed to observe. Three phenomena (trust asymmetry, the detection horizon, and the completion drive) are read here as empirical signatures of stratum 2 dynamics, observable in deployed systems and resistant to characterization at the model layer alone, and we offer them as evidence that the relational stratum is a tractable empirical object the field is already documenting in fragments. By relational architecture we name the deployment-layer infrastructure the field would need to treat the human-agent relationship as a first-class object, with relational evaluation as the corresponding measurement discipline. We close with a five-item research program for studying the relational stratum rigorously. This position would be falsified if the three phenomena we name are shown to be fully characterizable through model-layer evaluation alone, without reference to the longitudinal human-agent interaction.