RD-Agent

RD-Agent is Microsoft's open-source research agent for machine-learning engineering, Kaggle work and quant experiments.

RD-Agent is a Microsoft open-source agent that runs research-and-development loops for machine-learning engineering. The authors report it as the top-performing machine-learning engineering agent on MLE-bench in their published results, and they have a Kaggle scenario, a general data-science agent and R&D-Agent-Quant for trading research (NeurIPS 2025). LiteLLM is the default backend for many model providers. A web UI (rdagent serverui) shows traces in real time. Related papers include FT-Dojo / FT-Agent for autonomous fine-tuning (ICML 2026) and a technical report on the overall framework. A live demo and documentation are linked from the repository.

RD-Agent at a glance

Best for: Researcher, Software engineer, Finance
Runs on: open source, Web
Pricing: Free and open source; model API usage is billed by the provider
Released: April 2024
Category: AI research tool
Website: rdagent.azurewebsites.net

Key features

ML engineering agent evaluated on MLE-bench
Kaggle scenario and a general data-science agent
R&D-Agent-Quant for trading research
LiteLLM backend for many model providers
Web UI for traces and live interaction
Papers, a technical report and a live demo

Use it for

Run an MLE-bench-style model-engineering loop
Have the agent try a Kaggle-style tabular task
Watch an experiment trace in the web UI

Alternatives to RD-Agent

Paper2Agent — It enables direct interaction with research papers and automatic result reproduction.
OpenResearch — Run parallel research workflows locally or on remote compute.
LLM Wiki Research — Compiles parallel agent research into source-backed markdown knowledge bases.
rLLM — Enables reinforcement learning training for AI agents with minimal code changes.
AI Scientist v3 — Automates end-to-end scientific research from task descriptions to complete papers.
DiscoveryRSI — Achieves high actionable-reference rates and supports multi-agent evolutionary research search.
MiroThinker — It handles multi-step reasoning and complex research tasks with high stability.
Lune — Accesses trusted knowledge from top conference papers for AI agents.
Auto Deep Researcher — Automates experiment operations to reduce manual babysitting of training loops.
Arbor — It grows a hypothesis tree and tests changes on held-out data.
MiroFlow — Achieves top rankings on multiple agentic benchmarks with modular architecture.
Jevon Tane — It provides a rigorous, pre-registered framework for AI agents to conduct science together.

All RD-Agent alternatives