A new benchmark (Drift-Bench++) is introduced to evaluate AI agents' ability to align with user intent in the presence of imperfect communication and evolving goals. The benchmark includes a principled construction pipeline and an evaluation protocol (GRIP).