RD-Agent is an open-source framework from Microsoft that uses LLM agents to automate research and development work, especially the data-and-model loop at the heart of applied machine learning. Its premise is bold: let AI drive data-driven AI. The project is under very active development on GitHub, where it has earned more than 15,000 stars, with commits continuing through late 2026.
RD-Agent organizes R&D as a closed loop of proposing ideas, implementing them, running experiments, and learning from feedback. It ships several scenarios out of the box. The data science scenario tackles Kaggle-style problems end to end, from exploration to a submission-ready model. R&D-Agent-Quant applies the loop to quantitative finance, co-evolving predictive factors and models, and has been written up in its own research paper. Other scenarios cover autonomous LLM fine-tuning and reading academic papers to reproduce and extend their methods. A web UI lets you watch experiments and traces in real time, and the framework is model-agnostic, working with popular LLM APIs and local setups. Because it comes from Microsoft Research, the repository also collects a growing list of accepted papers, including work presented at NeurIPS, ICML, and ACL venues.
RD-Agent is completely free and open source under the MIT license. You supply your own LLM API keys or local models and your own compute for experiments; there are no subscriptions or hosted tiers.
RD-Agent is for ML researchers, data scientists, and quant teams who want to automate repetitive experimentation and explore ideas faster. It is also a reference implementation for anyone studying how autonomous agents can do genuine research-grade work rather than simple chat tasks.