Aligning Agentic AI: Dynamic Value Grounding in Open-Ended Environments

ABSTRACT

Traditional AI alignment techniques like RLHF are insufficient for Agentic AI, as autonomous systems frequently encounter novel edge cases in open-ended environments that were absent from their training distributions. We propose Dynamic Value Grounding (DVG), a framework where agents actively query a foundational 'Ethics Oracle' during multi-step execution to resolve moral ambiguity. By treating alignment as a continuous, runtime constraint rather than a static training objective, our agents demonstrate a 72% reduction in misaligned actions during highly unconstrained simulated resource-gathering tasks.