Multi-Agent Robotic Control with Onboard Vision-Language Models