Agentic Microphysics: Safety Research Must Shift from Individual Models to Agent Population Dynamics
arXiv·medium signal
This manifesto argues that as AI systems acquire planning, memory, tool use, and persistent identity, safety analysis can no longer focus on isolated models. Population-level risks emerge from structured interaction among agents through communication, observation, and mutual influence. The paper proposes 'agentic microphysics' — a methodology borrowed from social physics — for analyzing collective agent behavior over time, filling a critical gap in current safety research.