Multimodal Reinforcement Learning with Adaptive Verifier for AI Agents