Automatically generate and backtest quantitative trading signals by having the agent propose formulas, code them, run historical backtests, and refine the approach
Enter Kaggle machine learning competitions by letting the agent read the problem, propose solutions, write and debug code, and submit results autonomously
Fine-tune large language models without writing training recipes by hand, the agent designs experiments and runs them autonomously
Assign a powerful reasoning model to the research role and a cheaper model to the coding role to balance quality against cost
Requires API keys for one or more AI providers, GPU resources are needed for self-hosted LLM fine-tuning experiments.
RD-Agent is a project from Microsoft that uses AI to automate parts of research and development work, specifically in data science and machine learning. The core idea is that a lot of R&D work follows repeatable patterns: propose an idea, write code to test it, evaluate the results, and repeat. RD-Agent builds software agents that can carry out this cycle with minimal human involvement. The project targets several practical scenarios. In quantitative finance, it can automatically generate and test trading signals by proposing mathematical formulas, coding them up, running backtests against historical market data, and refining the approach based on results. In general data science competitions, it can enter Kaggle-style machine learning contests by reading the problem description, proposing solutions, writing and debugging code, and submitting results. It currently ranks at the top of MLE-bench, a public benchmark that measures how well AI agents handle machine learning engineering tasks drawn from 75 real Kaggle competitions. The system is built around a loop of two main roles: a research agent that proposes ideas and a development agent that implements them. Different AI models can be assigned to each role. For example, one configuration uses a more capable reasoning model for the research side and a faster, cheaper model for writing code on the development side. Installation is through the standard Python package manager. The project supports multiple AI providers through a compatibility layer, so you are not locked into a single service. A web interface is included for watching runs in real time and reviewing the agent's reasoning traces. The repository also covers fine-tuning large language models autonomously, where the agent designs training recipes and runs experiments without human intervention. Papers from this project have been accepted at major machine learning conferences including NeurIPS 2025 and ICML 2026. The full README is longer than what was shown.
← microsoft on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.