microsoft/RD-Agent

What is RD‑Agent?

RD‑Agent (R&D‑Agent) is an open‑source Python framework from Microsoft that lets large‑language‑model (LLM) agents automate end‑to‑end research‑and‑development pipelines. It bundles together two complementary agents:

• R‑agent (Research) – proposes ideas, reads papers or reports, designs new features or factor models.
• D‑agent (Development) – takes the proposals and implements them: writes code, runs experiments, trains models, and evaluates results.

The system is built around LLM‑driven reasoning (ChatCompletion, JSON‑mode, embeddings) and a multi‑agent orchestration loop that iterates until a target objective is met. It ships with a set of ready‑made “scenarios” that showcase the loop in concrete domains:

Scenario	Goal	Demo link
Quantitative‑Finance (RD‑Agent(Q))	Jointly optimise factor models and trading strategies using Microsoft QLib	Live demo
Data‑Science / Kaggle	Automatic feature engineering, model selection and submission for Kaggle competitions	Live demo
Research Copilot	Parse research papers or financial reports and turn them into runnable model code	Live demo
Medical Prediction	Build and iterate predictive models for clinical time‑series data	Live demo
FT‑Agent	Self‑supervised fine‑tuning of LLMs for a downstream benchmark	(see rdagent/app/finetune/llm)

Why it matters

• Benchmark‑first – The authors evaluate RD‑Agent on MLE‑bench (75 Kaggle‑style tasks) and on a new Agent² RL‑Bench that measures an agent’s ability to engineer post‑training RL pipelines. RD‑Agent currently holds the top spot on MLE‑bench.
• Modular, extensible design – New domains can be added by writing a small “scenario” module that defines the state, actions, and evaluation metric. The core loop stays the same.
• Pluggable LLM back‑ends – Supports any model reachable through LiteLLM (OpenAI, Azure OpenAI, DeepSeek, SiliconFlow, etc.) and can be swapped via a simple .env file.
• Open‑source & reproducible – Distributed on PyPI (pip install rdagent), with Docker support, CI badges, type‑checking (mypy), linting (ruff), and full documentation.

Quick start (Linux only)

# 1. Create a conda env
conda create -n rdagent python=3.10 && conda activate rdagent

# 2. Install the package
pip install rdagent   # or `make dev` from source

# 3. Configure LLM credentials (example for OpenAI)
cat <<EOF > .env
CHAT_MODEL=gpt-4o
EMBEDDING_MODEL=text-embedding-3-small
OPENAI_API_KEY=YOUR_KEY
OPENAI_API_BASE=https://api.openai.com/v1
EOF

# 4. Verify the setup
rdagent health_check

# 5. Run a demo – e.g., quantitative‑trading loop
rdagent fin_quant

The command launches a background Docker container (for Qlib) and starts the R‑/D‑agent loop. Results can be inspected in a web UI (rdagent server_ui) that visualises each iteration’s proposal, code, and performance metrics.

Where to learn more

• Live demo & videos – https://rdagent.azurewebsites.net (factor loop, model loop, research copilot, etc.)
• Documentation – https://rdagent.readthedocs.io/en/latest/
• Technical report – https://aka.ms/RD-Agent-Tech-Report (arXiv 2505.14738)
• Benchmarks – MLE‑bench leaderboard and Agent² RL‑Bench paper (arXiv 2604.10547)
• Community – Discord (invite link in README) and a WeChat group for Chinese‑language users.

---

TL;DR: RD‑Agent is a Microsoft‑maintained, Linux‑only Python library that lets LLMs act as autonomous R&D assistants. It already beats other agents on a large ML‑engineering benchmark and ships with ready‑to‑run scenarios for finance, data science, and research‑paper automation.