Ask WhAI:Probing Belief Formation in Role-Primed LLM Agents