.png)
Course description
Most researchers who use AI for coding do it through a chatbot: paste a snippet in, copy an answer out. Coding agents work differently. They read your whole project, plan multi-step changes, run your scripts, read the error messages, and revise until the result meets a definition of “done” that you set. This workshop starts by making that difference concrete, and then shows what it means for everyday research work.
The workshop is editor- and model-agnostic: participants are encouraged to use whatever they are comfortable with and what is approved for their data, and we discuss the pros and cons of each option. We demonstrate VS Code (the most flexible editor) and Positron (the most familiar for R and RStudio users), show OpenCode briefly, and list the alternatives, including the Claude and ChatGPT/Codex desktop apps.
We combine these with the GPT UiO API, which offers models approved for green, yellow, and even red data, and with personal Claude and ChatGPT subscriptions, which are only appropriate on a personal device with green data. Local models are mentioned but are beyond the scope of this workshop. We also show how to steer and extend an agent, within UiO’s guidelines, with project instruction files (CLAUDE.md / AGENTS.md), skills, and connectors (MCP servers).
The workshop does not depend on one programming language: we cover in depth which combinations of language (R, Stata, Python, SAS, MATLAB, and others), editor, model, and access route work. Researchers who use point-and-click tools such as SPSS, GraphPad, or JMP are welcome too: we show how an agent can help you move an analysis you already know into code. We start by finding out who is in the room (disciplines, methods, academic level, and statistical software) and adapt the exercises and demonstrations accordingly; for example, the TSD part is shortened if no one works with sensitive data. Exercises may include cleaning a dataset, writing and documenting an analysis script, translating code between languages, and debugging a statistical model.
Three topics get particular attention because they decide whether agentic coding is safe to use in research:
- Git, GitHub, and DOI-issuing archives such as Zenodo, taught pragmatically and only to the extent necessary for researchers. They make every change the agent proposes transparent, reproducible and reversible, and let you share your code and archive each released version with a DOI so that it can be cited in peer-reviewed publications.
- Sensitive data. “Red” and definitely “black” data, such as health registry data, must remain in TSD, and participants do not bring any real data. Using a Norwegian registry as the example, we demonstrate how to compute population characteristics (mean, median, SD/SE, confidence intervals) from red or black data inside TSD, generate simulated data from these and the registry’s codebook, develop and test the analysis code locally with an agent, and then import the finished code into TSD to run on the real data. We also cover what to watch out for when exporting summary statistics from TSD (disclosure control, small cells, extreme values).
- Responsible use. Whether agentic coding is a good idea for a given research project, and what can go wrong: an agent can write code that runs without errors but is conceptually or statistically wrong. We discuss how much subject and statistical knowledge a researcher should already have before relying on an agent, and how its use should be disclosed in manuscripts and to journals.
The workshop is for researchers rather than software developers, at a beginner to intermediate level in coding, and in particular for those who want to produce reproducible code that can be shared and cited alongside their publications.
It does not teach statistics or subject-matter knowledge: participants should already analyze their own data as part of their current role. It focuses on project and agentic AI setup to increase productivity in reproducible and ethical ways including checking the agent’s output and data privacy. Participants leave with a working setup, a project template, and a complete agent session behind them.
Target audience
Researchers at all career levels and fields, from master’s students and PhD candidates to postdocs and associate and full professors, as well as research assistants and research support staff (e.g., data managers, research engineers, statisticians), who analyze data as part of their current role, whether in code or in point-and-click software, and are at a beginner to intermediate level in agentic coding, Git, and coding in general.
Fields: any field that does quantitative or computational data analysis, such as the life and health sciences (including medicine), the social sciences (economics, psychology, political science, sociology), and the natural sciences and engineering (physics, chemistry, geosciences, and others).
The workshop is relevant both for those whose data can be analyzed on their own computer and for those who work with sensitive or personal data in TSD.
Learning outcomes
After this workshop, participants will be able to:
- Explain how a coding agent differs from autocomplete and from chat-based AI, and when each one is the right tool
- Choose and set up a combination of editor, model, and access route that fits their language, their data, and UiO’s guidelines, and know the pros and cons of the alternatives
- Put a research project under version control with Git and GitHub, organize the project folder, and archive released versions with a citable DOI (e.g., via Zenodo or a similar archive)
- Write a project instruction file (CLAUDE.md / AGENTS.md) that tells the agent the conventions and the definition of “done,” and extend the agent with skills and connectors (MCP servers) within UiO’s guidelines
- Use an agent for typical research tasks: data cleaning, analysis scripts, documentation, debugging, translating code between languages, and moving a point-and-click analysis into code
- Generate simulated data from a codebook and summary statistics (e.g., computed in TSD), develop code against it, and bring the validated code into TSD, in line with UiO’s rules on data classification and AI use
- Check the agent’s output critically (tests, sanity checks, reading the diff) and take responsibility for the results, judge when agentic coding is appropriate given their own subject and statistical knowledge, and disclose its use in publications
Prerequisites
- Entry level: beginner to intermediate (in agentic coding, Git, and coding in general).
- Participants must analyze their own data as part of their current role, in any tool. The workshop does not teach statistics or subject-matter knowledge: participants are expected to know which analyses they need and how to interpret them. Users of point-and-click software (SPSS, GraphPad, JMP) are welcome; no programming or syntax experience is required. No experience with Git, GitHub, the command line, or AI coding tools is needed.
- Preparation is required: participants receive a setup guide in advance and are expected to work through it beforehand (install an editor, Git, and their usual statistical software, or R or Python for point-and-click users, create a GitHub account, and set up access to an agent). The first part of the day is reserved for fixing remaining setup problems, and participants whose setup works are asked to help their neighbors.
Required materials
- A laptop (Windows, macOS, or Linux). On UiO-managed machines, VS Code and Git are available in Company Portal / Software Center, and Positron installs without admin rights.
- An editor of your choice (e.g., VS Code or Positron), Git, and the statistical software you normally use (R, Stata, etc.); point-and-click users install R or Python instead, as described in the setup guide.
- A free GitHub account and, ideally, the GitHub CLI installed beforehand (this may require administrator rights).
- Access to an AI model: the GPT UiO API (participants apply for access themselves; the setup guide explains how) or, on a personal device and for green data only, a personal Claude or ChatGPT subscription. Low-cost open-weight coding plans are discussed but not recommended.
- Internet access.
- Participants work only with simulated or public data and must not bring real sensitive or personal data; the TSD demonstration is done by the instructor.
Lunch
- There will be a complementary buffet-lunch at 12:00-13:00 from the Sverdrups spiseri (Georg Sverdrups Hus, 2nd floor).
- After you get your lunch, join us at Store møterom where we have reserved seats for #DSDays2027 participants.
Number of participants
30
Contact
If you have any questions regarding this course, or anything else about the DSDays 2027 event, please contact us per email: digitalscholarship@ub.uio.no