Agentic Multimodal Large Language Models (MLLMs) are reshaping scientific discovery. They do more than answer questions: they read multimodal evidence, reason over scientific constraints, use tools, and close parts of the research loop. This survey explains how such systems are built for scientific discovery, moving from multimodal perception to training, inference, evaluation, and human-AI collaboration. Across this pipeline, scientific agents face bottlenecks that generic assistants often hide, including scarce expert annotations, evolving domain structures, complex reasoning, immature tool interfaces, safety and toxicity risks, and unfaithful results. We highlight design choices for turning agentic MLLMs into trustworthy, evidence-grounded collaborators. Curated resources are available at https://github.com/JianJinglin/awesome-agentic-AIScientists.