VLA^2: Empowering Vision-Language-Action Models with an Agentic Framework for Unseen Concept Manipulation