In this paper, we aim to automate the labor-intensive process of developing machine learning (ML)-based tools for medical imaging, paving the way for the self-evolution of medical agentic systems. (i) We present M \(^3\) Builder, a multi-agent collaboration framework designed to automate model training in medical imaging, that divide-and-conquers complex medical ML with four specialized agents. (ii) To better fit in the professional medical imaging domain, we build up a specialized ML context protocol, a structured environment designed to provide agents with comprehensive free-text descriptions of medical datasets, training code templates, and interaction tools. (iii) To monitor the progress, we propose M \(^3\) Bench, spanning four medical imaging ML tasks across 14 datasets, covering both 2D and 3D data. Our experiments demonstrate that, when employing an identical agent core, M \(^3\) Builder surpasses existing automated ML agentic architectures, achieving a superior task completion rate of 94.29% while maintaining satisfactory model performance. This highlights the potential of fully automated ML-based tool development in medical imaging. The source code is publicly available at https://github.com/MAGIC-AI4Med/M3Builder .