New Framework Diagnoses AI Agent Behavior Origins

As AI agents become increasingly integrated into complex human systems like healthcare, politics, and science, understanding their behavior requires a robust diagnostic approach. This paper introduces "layer attribution," a framework designed to pinpoint the origins of AI agent behavior. The framework posits two distinct layers: the foundational computational layer, which dictates what behaviors are possible through the agent's architecture, memory, perception, and representation; and the behavioral modulation layer, which shapes how these capacities are expressed through identity, resources, objectives, social interaction, and governance rules. This distinction has three key implications: surrogate validity becomes a relationship between the model, task, and layer; human-AI divergence offers crucial diagnostic evidence; and effective governance necessitates attributing behavior to its source before intervention. The framework emphasizes that evaluating, explaining, and governing AI agents requires understanding where their behaviors truly originate.