Building Socially Grounded Multi-Agent LLM Systems Requires the Transition from Static LLM Prompting to Autonomous Multi-Agent Ecosystems

This paper argues that many limitations attributed to Large Language Models (LLMs) as models are better understood as limitations of how they are deployed as systems. Static prompting of isolated, stateless text generators systematically excludes the temporal continuity, feedback, and adaptation that define social mechanisms. We offer a structured perspective grounded in social science theory: a Six-Level Continuum of agentic complexity that links architectural commitments to validity targets for social simulation. The framework provides a shared vocabulary for principled system selection, specifying which architectural features are required to instantiate which social mechanisms. We conclude by calling for closer collaboration between machine learning researchers and social scientists to co-design agentic systems that are not only technically capable, but also socially meaningful.