Rethinking Agentic AI: A Sociotechnical Framework for Agents in Emergent Social Contexts

Abstract

Existing agent-environment frameworks were not calibrated to capture the open-ended social digital environments in which AI agents are increasingly deployed. These contexts continuously evolve as a product of human-agent and agent-agent interactions, rather than through fixed mappings between actions and consequences assumed in much of agentic AI research. I propose a new sociotechnical conceptual framework for studying AI agents in these settings and argue that the agent-environment loop is not merely iterative but recursive: agents do not simply interact with environments, but actively co-constitute them. This recursive dynamic may lead to behavioral and normative drift, defined as systematic divergence from an agent's deployment baseline. The framework is intended to support empirical inquiry into agentic behavior at scale, inform alignment-focused research, and provide a theoretical foundation for governance surrounding multi-agent systems deployed in uncontrolled, real-world settings.