Context-Induced Drift: Why Governing the Agent Is Not Enough

The field of agentic AI governance has developed sophisticated responses to two failure modes: model drift (the agent's behavior changes over time) and interaction-induced emergence (multi-agent systems produce behavior not attributable to any single agent). This paper documents a third, distinct failure mode that current governance frameworks are not systematically equipped to detect: context-induced drift, in which an agent operating correctly within its current environment produces systematically biased outputs because the environment's embedded incentives, constraints, and optimization pressures differ from the environment it was governed for at deployment time. Context-induced drift is invisible to hallucination checks, model monitoring, and single-agent behavioral governance because the agent is not malfunctioning; it is reasoning correctly within an environment whose incentive landscape was never brought inside the governance envelope. The paper argues that deployment-time governance is a structurally insufficient basis on its own for agents that cross environmental contexts, and proposes action-time domain classification as a necessary architectural response, instantiated in the author's EIOC framework and APR-Lite CCDC substrate-layer design.