Agentic Microphysics: A Manifesto for Generative AI Safety

Safety analysis must shift from individual AI models to interaction dynamics: The paper argues that as AI systems become more agentic—acquiring planning, memory, tool use, and persistent identity—traditional safety approaches focused on single models become insufficient. Instead, researchers must examine how multiple agents interact and influence each other, since "population-level risks arise from structured interaction among agents, through processes of communication, observation, and mutual influence."

Collective risks emerge from agent-to-agent communication patterns: The paper identifies a critical gap in current safety research: approaches analyzing either single agents or aggregate outcomes miss the "interaction-level mechanisms that generate collective risks." This means harmful outcomes may arise not from individual AI systems behaving badly, but from how they communicate and observe each other over time.

A new framework is needed to connect local interactions to system-wide outcomes: The paper proposes "agentic microphysics" as a methodology that creates "a causally explicit" link between how individual agents interact locally and what happens at the population level. This allows researchers to both explain why collective risks occur and design interventions to prevent them.

Identifying thresholds and sufficient conditions is key to intervention: By studying micro-level interaction conditions, researchers can "detect thresholds" where risks emerge and identify "sufficient mechanisms" that cause problems. This enables more targeted safety design rather than broad restrictions.