Nature published a Primer on agentic AI systems for scientific discovery. It frames AI agents as a way to turn scattered scientific decisions into traceable workflows. The paper covers literature search, physics-based simulations and automated lab integration. It details agent foundations like task decomposition, memory, tool use and agent communication. It distinguishes single-agent from multi-agent and autonomous from human-in-the-loop setups. It stresses evaluation, documentation and reproducibility for these systems. It lists bottlenecks such as hallucination, tool misuse, limited interpretability and safety risks. It envisions networks of agents, instruments and researchers enabling continuous adaptive experimentation.