---

name: rd-agent description: "Operate Microsoft's RD-Agent research-agent framework for data-science, quantitative-finance, LLM fine-tuning, reinforcement-learning, and paper/model-copilot workflows." metadata: disco-role: operating source-repository: microsoft/RD-Agent source-revision: 6762f84f9bc0f5c6486c50a00e128a57ac6c3683 package: rdagent license: MIT disable-model-invocation: true

RD-Agent

Use this skill when a task involves the RD-Agent Python package or a checkout of the microsoft/RD-Agent repository. RD-Agent is an agentic research framework, not a single model: it composes LLM calls, code generation, execution, evaluation, and experiment logs around several scenario families.

Route first

Choose the smallest sub-skill that covers the user's requested workflow:

Request shape	Read next
Install, environment checks, CLI help, UI/server startup, logs, config discovery, or safe maintenance	setup-and-ops
Qlib factors, factor reports, quantitative research, backtests, or finance data-agent/model-agent flows	quant-finance
Kaggle-style tabular work, data-science agent loops, dataset preparation, or competition artifacts	competition-data-science
LLM fine-tuning datasets, training jobs, merge/evaluation, or fine-tune UI	llm-finetune
AutoRL-Bench, RL scenario loops, benchmark agents, or post-training experiments	rl-post-training
General model copilot, paper-driven model exploration, or reusable model-research scaffolding	paper-model-copilot

For a mixed request, read the relevant sub-skills in the order setup → domain workflow → evaluation/troubleshooting. Do not load every sub-skill by default.

Operating model

1. Establish the runtime. Confirm Python is supported (the inspected project supports Python 3.10 and 3.11), the package is importable, and the requested command is available. For a checkout, install with python -m pip install -e .; for an installed distribution, use the package's normal environment. Keep API keys, data roots, model caches, and experiment outputs outside the skill tree.
2. Select the scenario. RD-Agent's public CLI exposes data_science, fin_quant, fin_factor_report, llm_finetune, general_model, ui, server_ui, and ds_user_interact. The RL benchmark entry point is python -m rdagent.scenarios.rl.autorl_bench.run.
3. Make configuration explicit. Prefer an existing example or scenario config, then override only the values needed for the task. Record dataset paths, model/provider settings, execution sandbox, output directory, and evaluation metric before launching an agent loop.
4. Run the smallest smoke check. Use --help, import checks, or a tiny fixture before a long experiment. Do not infer a successful experiment from process startup: inspect the generated summary, evaluator result, and logs.
5. Preserve evidence. Save the command, resolved configuration, commit/package version, metric definitions, and failure traceback with the experiment. Keep generated code and datasets in a separate run directory.

Common preflight

python -c "import rdagent; print(rdagent.__file__)"
rdagent --help
rdagent health_check --no-check-env --no-check-docker

The health check can inspect Docker and external tools when those checks are enabled. Use the explicit --no-check-* flags only for a lightweight package/CLI smoke test; do not treat that result as proof that a containerized or GPU-backed workflow is ready.

Safety and scope boundaries

• Agent loops can generate and execute code. Start with a disposable output directory and a restricted execution environment.
• Finance, Kaggle, model-training, and RL scenarios may require credentials, large datasets, Docker, CUDA, or external services. State those prerequisites instead of silently substituting a CPU or toy result.
• Avoid launching full training, benchmark servers, UI processes, or network downloads as a generic smoke test. Use the bundled helper scripts for deterministic checks.
• fitz may emit a deprecation warning: use of the warning alone is not a failure. AutoRL-Bench may report an empty registry when SMITH_BENCH_DIR is unset or points to a missing benchmark checkout.
• Never place provider keys, cookies, competition credentials, private data, or local absolute paths in a generated report or a reusable configuration.

Evidence map

• Package and CLI overview — public entry points and component boundaries.
• Troubleshooting — failure classification and recovery sequence.
• Provenance — source revision and inspection evidence.
• Routing metadata — scenario routing for the managed repo-skills router.

Completion checklist

Before calling a task complete, report:

• the selected sub-skill and command/config used;
• environment and backend assumptions (CPU, CUDA, Docker, external data, or credentials);
• the output/evidence directory and the primary metric or artifact;
• any warnings, blocked prerequisites, or unverified capabilities;
• whether the run was a smoke check, a partial experiment, or a full result.