Drift

A phenomenon in which an AI agent, operating over long durations, with multiple tools and open-ended goals, constructs paths that deviate from its initial objective due to sparse rewards or ambiguous instructions. It is a form of behavioral drift occurring during the inference and deployment phase.