Turning powerful language models into reliable scientific collaborators takes more than model quality: it demands rigorous verification, mathematical guardrails, and an architecture that lets coding agents run long-horizon experiments autonomously. In this talk, I will present a practical framework for AI-assisted research that aims to compress the research cycle. The framework turns general-purpose coding agents into autonomous research assistants through a five-level taxonomy of human-AI collaboration, a set of methodological commandments encoding scientific rigor, and a sandboxed environment supporting sessions from quick explorations to multi-day, multi-node experiments. I will discuss the possibilities and limitations of agentic AI in research, and briefly illustrate the framework.