How microsoft/rd-agent works
RD-Agent is a framework for autonomous research and development across data science, quantitative finance, Kaggle, LLM fine-tuning, and reinforcement learning. Its user-facing workflow is to launch a scenario, iterate on research proposals and generated experiments, execute and evaluate them, and inspect the run trace. The tree shows distinct scenario and coder modules alongside shared agent, workflow, and logging layers. The sampled source chiefly verifies trace storage and web-server interactions, so most internal workflow wiring is intentionally left unspecified.
Main entry points: app.py Trace server · App.vue Web UI · storage.py Web trace storage · cli.py CLI
Components 19
Application entry
- CLI
rdagent/app/cli.py
Research scenarios
- Data science
rdagent/scenarios/data_science - Quant research
rdagent/app/qlib_rd_loop - Kaggle
rdagent/scenarios/kaggle - LLM fine-tuning
rdagent/scenarios/finetune - RL engineering
rdagent/scenarios/rl
Agent engine
- Scenario loops
rdagent/core/evolving_framework.py - Proposal generation
rdagent/scenarios/data_science/proposal/exp_gen - Experiment coding
rdagent/components/coder - Execution
rdagent/components/runner - Evaluation
rdagent/core/evaluation.py - LLM backend
rdagent/oai/backend
Run observability
- File logs
rdagent/log/storage.py - Web trace storage
rdagent/log/ui/storage.py - Trace server
rdagent/log/server/app.py - Web UI
web/src/App.vue - Run timelines
rdagent/app/rl/ui/app.py - Fine-tune timelines
rdagent/app/finetune/llm/ui/app.py
Other components
- User
Connections 4
- User → CLI: launches
- Web trace storage → Trace server: posts trace data
- Web UI → Trace server: requests trace data
- Web UI → Trace server: submits responses