RD-Agent is Microsoft's open-source framework for automating research and development workflows with LLM agents. After 30 days of using it to run data experiments for saas.pet's tool analysis, here is the real story on the propose-execute-feedback loop, what it automates well, and why it is not a magic research button.
RD-Agent is the closest thing I have seen to an actual research assistant that does work, not just chat. It is Microsoft's framework for automating R&D: you give it a research goal, it proposes an approach, executes experiments, reads the results, and iterates. I use it for saas.pet's data analysis, things like exploring patterns in our tool database or testing whether a new categorization rule improves search quality. The propose-execute-feedback loop is real: it ran 40 iterations of an experiment overnight and came back with a result I could actually use. The honest caveats: it is designed for data science and ML workflows, not general software development, and it needs a well-defined environment to run experiments in. It is also heavy, Docker and a GPU recommendation for serious runs. For anyone doing repetitive data experiments, it is genuinely useful. For general coding tasks, use a coding agent instead.
What RD-Agent actually is
RD-Agent is an open-source framework from Microsoft Research for automating research and development workflows using LLM agents. It is built around a propose-execute-feedback loop: the agent proposes a hypothesis or approach, executes it in the environment, reads the feedback from the results, and iterates. It is designed for data science and ML research workflows, covering things like feature engineering, model selection, experiment running, and report generation. As of August 2026 the repo has 14,326 stars, an MIT license, and while the latest tagged release is v0.8.0 from November 2025, the codebase has commits from August 2026, so it is actively developed. The project describes itself as addressing the core question of how LLMs can enhance industrial R&D productivity, and its design reflects that: it is a serious research tool, not a chatbot wrapper.
Why I tried it: saas.pet's data experiments
saas.pet maintains a database of AI tools with categories, review scores, and usage data, and I kept having recurring questions: which category features predict high engagement, does a new tagging rule improve search relevance, what patterns exist in review lengths and scores. Each of these used to mean writing a script, running it, looking at output, adjusting, and repeating. RD-Agent promised to automate exactly that loop, so I set it up in a Docker container with a small experiment environment pointing at a copy of the tool data. The first run was a categorization experiment: I asked it to test whether adding a domain field improves search result relevance, and it proposed a metric, ran the experiment, measured the baseline, tried variants, and came back with a comparison table. It was not a perfect research assistant on day one, but the loop was real and the output was usable.
The propose-execute-feedback loop in practice
The core mechanic is a loop: propose, execute, feedback, iterate. In practice this means RD-Agent does not just answer a question, it runs actual experiments. In my categorization test, it proposed a metric for measuring relevance, wrote the evaluation script, executed it against the data, read the results, identified that the new field helped for some categories and hurt for others, and iterated on the approach. The most impressive run was an overnight experiment where it executed 40 iterations of a feature selection task and left me a summary of what worked and what did not. The feedback mechanism is what makes it different from asking an LLM directly: the agent sees real execution results, not guesses. It also generates reports and tracks experiments, so you can see the history of what was tried. For anyone whose work involves repetitive data experiments, this loop is the product.
What it automates well and what it does not
RD-Agent is strongest at structured data science workflows: feature engineering, model evaluation, experiment comparison, and report generation. It handles the boring parts of research, the running of experiments and the summarizing of results, and it is good at that. It is not a general software development agent: it will not build a web app or fix a production bug, that is what coding agents like Cline or OpenHands are for. It is also not a replacement for domain expertise, you still need to define the research question and validate the results. The sweet spot is: you have a data experiment you would run manually, RD-Agent runs it and iterates while you do something else. The setup expectation matters too: it wants a defined environment, Docker is the recommended path, and serious ML runs benefit from a GPU. On a CPU-only machine, small experiments work, large model training does not.
Comparing with OpenHands, Cline, and plain notebooks
Against coding agents like OpenHands and Cline, RD-Agent is a different category: those are built for software engineering tasks, writing and fixing code across a repo, while RD-Agent is built for research workflows, proposing and running experiments. Using RD-Agent to write a web feature would be awkward, and using Cline to run a 40-iteration feature selection study would be equally awkward. Against plain Jupyter notebooks, the comparison is about automation: a notebook is a manual loop where you write, run, read, adjust, while RD-Agent closes that loop automatically. For a solo operator like me, the time savings come from the loop running unattended, especially overnight. The honest middle ground is that RD-Agent is one tool in a stack, not the whole stack: it handles the experiment loop, and you still need coding agents for software work.
Pricing and setup cost
RD-Agent is MIT licensed and free. The setup cost is real: it runs in Docker, needs Python dependencies, and the recommended path includes a fair amount of configuration for the experiment environment. I spent an afternoon getting the first end-to-end run working, including the Docker setup and pointing it at the data. The model cost depends on what you point it at: it supports OpenAI-compatible APIs, so DeepSeek works and keeps costs low, my overnight 40-iteration run cost under a dollar. There is also support for local models, which matters if you want the experiment data to stay on your machine. The resource cost is the bigger consideration: a serious ML experiment wants a GPU, and CPU-only runs are limited to small datasets and simple models. For data-scale work, factor in the compute cost alongside the model cost.
Limitations and final verdict
The honest limitations. First, it is specialized: data science and ML research workflows only, general development is out of scope. Second, the environment setup is heavy, Docker, dependencies, and a well-defined experiment space, and the learning curve is steeper than a chatbot or a notebook. Third, the release cadence is odd: the latest tag is from November 2025 while commits continue, which can confuse version pinning. Fourth, results depend on your research question being well-defined, garbage questions produce garbage experiments. Who should skip RD-Agent: anyone whose work is software development rather than data experiments, and anyone who will not invest in the setup. For data practitioners doing repetitive experiments, it is a 4 out of 5, the loop is genuinely useful and the overnight automation is the killer feature. It will not replace your judgment, but it will run your experiments while you sleep.
What I liked
- Real propose-execute-feedback loop, not just chat
- Runs experiments unattended, overnight automation works
- Covers feature engineering, model selection, report generation
- MIT license, Microsoft-backed, actively developed
- OpenAI-compatible model support, cheap with DeepSeek
- Tracks experiment history and generates reports
What I did not like
- Specialized for data science and ML, not general development
- Heavy setup: Docker, dependencies, defined environment
- GPU recommended for serious ML runs
- Latest tag is old (v0.8.0, Nov 2025) despite active commits
- Results depend on well-defined research questions
Originally published on https://saas.pet/reviews/rd-agent-review
Top comments (0)