The paper presents R&D-Agent, a novel framework designed to address the growing complexity and expertise requirements in machine learning engineering (MLE) by formalizing the MLE process into a structured, autonomous system. It decouples the workflow into two phases and six components, transforming agent design from an ad-hoc, craftsmanship-based approach into a principled and testable methodology. This framework aims to overcome the labor-intensive and iterative nature of high-level MLE tasks, which are not fully alleviated by current crowd-sourcing platforms, by enabling the development of efficient agents inspired by human experts.

R&D-Agent demonstrates significant practical impact by achieving state-of-the-art performance, with its built agent ranking as the top-performing MLE agent on the MLE-Bench benchmark, attaining a 35.1% any medal rate. This highlights the framework's ability to speed up innovation and improve accuracy across diverse data science applications. The authors have open-sourced the framework on GitHub, promoting further research and development in autonomous data science, while showing that existing agents can be viewed as partial optimizations within this comprehensive framework.