Intro to Agentic Coding

The Intro to Agentic Coding workshop covers the working principles behind effective and safe agentic coding for researchers. AI coding agents (Claude Code, GitHub Copilot, Codex, Cursor, OpenCode, and their successors) can read a repository, run commands, edit multiple files, and revise their own output. They shorten the time between an idea and a result, and they shorten the time between a mistake and a misleading result by the same amount. This lesson is about principles rather than one product: limiting access, planning, specifying, verifying, and managing cost apply unchanged across tools. Where the mechanics differ, episodes give the equivalents for Claude Code and GitHub Copilot.

Three principles run through the lesson:

1. Stay in the driver’s seat. The agent writes the code. You review it, and you decide what is merged to main.
2. Work feature by feature, not project by project. A feature is one thing you can check. An underspecified prompt does not produce random code; it produces the statistically typical code for the average project, and research data is rarely the average case.
3. Assume nothing; verify everything. Rely on test-driven development. Good data science practice still applies in full: examine the data, know the distributions, and know what the model is responding to. Code that runs without error and scores well can still be wrong.

NotePre-alpha lesson: evolving as new info and policies come out

This lesson was developed by Chris Endemann, Tracy Reuter, and Tejvir Mann as part of the 2026 Machine Learning Marathon (MLM26), where it was delivered as a two-hour “Agentic Coding: (Developing) Best Practices” workshop. It is in the Carpentries pre-alpha phase: the materials are in early development and will keep changing as tools, institutional policies, and community best practices evolve. Have feedback or suggestions? Please post an issue on the lesson’s GitHub repository.

Topics covered

1. What is agentic coding? How agents differ from chat assistants and autocomplete, and what they can access.
2. Safety and security. Limit what the agent can access: web routes and dev containers, no agents on machines holding restricted data, and keeping API keys in a password manager rather than a file.
3. Trust. Packages, models, and providers: supply-chain risks and how to vet what your agent installs and calls.
4. Planning with agents. Repository reconnaissance and plan mode before any code is written.
5. Feature-based development and good prompting. Specify one checkable feature at a time.
6. Verification and testing. No escaping good data science: tests, assertions, CI, and reviewing agent-generated analyses.
7. MCP tools and skills. Extending your agent with the Model Context Protocol (MCP) and reusable skills, including a hands-on GitHub MCP exercise.
8. Cost, context, and energy. Managing token usage, context windows, and environmental footprint.
9. Documentation. Notes to your future self (and your agent).
10. What the research shows, and where this leaves us. Evidence on agentic coding productivity and quality, and an outlook.

Prerequisites

Learners should have:

• Comfort working at a command line and with git basics (clone, branch, commit, push).
• Some experience writing Python and doing basic data analysis (pandas, scikit-learn, or similar). No local Python installation is needed; the agent’s sandbox runs the code.
• Access to at least one agentic coding tool. The setup page lists options, including free tiers (Claude Code, GitHub Copilot’s education tier, and the open-source OpenCode).
• Ideally, a small project of your own to point the agent at. The exercises work on your repository, and a fallback starter is provided where it matters.

WarningNo agentic tools on machines holding sensitive or restricted data

The lesson’s setup page recommends a web-based route (the agent works on a cloud copy of a GitHub repository with no access to your machine) and requires a dev container for any locally run agent. If a machine stores restricted data (FERPA, HIPAA/PHI, CUI, export-controlled data, or anything under a data-use agreement), do not install or run any agentic tool on it. UW-Madison users should also review their data-privacy settings; see it.wisc.edu/ai.

Estimated time to complete

The workshop version runs about two hours and presents the main points. The full lesson (all ten episodes) is the extended reference and takes longer to work through independently. The instructor notes map workshop blocks onto episodes.

Register to take this workshop in Madison!

This workshop is part of UW-Madison’s local Carpentries community efforts to develop advanced ML/AI educational materials. To stay updated about upcoming workshops, subscribe to the Data Science @ UW Newsletter.

Alternatively, work through the materials independently!

The full lesson is available as open-source materials (CC-BY 4.0). Visit the lesson materials to explore on your own, or read all episodes on one page. If you’re at UW-Madison, join Coding Meetup (Tue/Thur, 2:30-4:30pm) for assistance.

Related resources

• Guide: Claude Code Best Practices: A deeper dive into permissions, context management, cost, and data privacy, using Claude Code as the primary lens.
• Guide: Claude Code Cloud Setup (Vertex AI & Bedrock): Route Claude Code through your UW-Madison institutional GCP or AWS account.
• Video: Intro to Git and GitHub: Brush up on the git basics (clone, branch, commit, push) that the exercises assume.
• Workshop: Intro to Machine Learning with Sklearn: The scikit-learn fundamentals behind the lesson’s starter project and verification exercises.
• Project: Machine Learning Marathon: The MLM26 challenges this workshop was developed to support.

Comments