Most attempts to automate research collapse the entire loop into one prompt-and-pray agent. RD-Agent's bet is that proposing ideas and implementing them are different skills, so it splits the work between an R agent that generates and ranks hypotheses and a D agent that turns the winning idea into working, tested code — then feeds the results back so each round is informed by the last.
What Sets It Apart
- Evolution over one-shot generation: the R↔D loop keeps refining factors and models across iterations, so quality compounds instead of riding on a single lucky completion.
- Domain depth, not toy demos: it ships concrete scenarios for quant trading (deep Qlib integration for factor and model co-evolution), Kaggle-style data science, paper-to-model reproduction, and LLM fine-tuning via FT-Agent.
- Numbers that hold up: on MLE-Bench the o3(R)+GPT-4.1(D) pairing reaches a ~30% medal rate (about 51% on low-complexity tasks) versus ~17% for AIDE o1-preview, and the quant variant reports roughly 2x the annualized return of benchmark factor libraries while using ~70% fewer factors.
Who It's For
Great fit if you have a well-scoped, metric-driven R&D loop — quant factor search, Kaggle pipelines, repeatable model experiments — where success is measurable and the agent can learn from each run. Look elsewhere if your work is open-ended or hard to score automatically: the framework leans on strong frontier models and a clean evaluation signal, and running long evolution loops burns real token and compute budget.
