Agentic Moderation: Multi-Agent Design for Safer Vision-Language Models