Video: best practices for AI-assisted coding with GitHub, tests and CI

This video covers the same material as my seminar “Best Practices for AI-Assisted Coding: GitHub, Testing, and Continuous Integration” (November 6, 2026), recorded from scratch for YouTube. It shows, on a small real project, how to work with an AI coding agent without losing control of your code: an AGENTS.md file and permissions, precise prompts and plan mode, tests, continuous integration, and subagents with separate roles. The demos use OpenCode with the free models of the National Research Platform (NRP), but the practices apply to any agent: Claude Code, Codex, Copilot CLI.

Watch it on YouTube: youtu.be/xEojDVgHsHw. Long agent waits are cut, and a card shows how much real time passed.

Set up the agent first

To follow along on your own machine, start from the companion post Set up an AI coding agent with free NRP models: NRP API key, OpenCode installation and configuration, permission rules and which model to pick.

The demo repository

Every demo runs on the same small Python package, which reads weather station data and computes daily statistics: github.com/zonca/weather-agent-demo. Each demo starts from a tag and ends at the next one, so you can check out any step and try it yourself:

git clone https://github.com/zonca/weather-agent-demo
cd weather-agent-demo
git checkout checkpoint-2   # the state at the start of demo 3

Demo	Starts from	Ends at
1. The agent learns the project	checkpoint-0	checkpoint-1
2. Asking and steering	checkpoint-1	checkpoint-2
3. Tests catch what you would miss	checkpoint-2	checkpoint-3
4. Continuous integration	checkpoint-3	checkpoint-4
5. Subagents keep the tutorial working	checkpoint-4	checkpoint-5

What is in the video

• 0:00 Introduction: why version control, tests and CI matter more, not less, when an agent writes the code
• 3:29 Keeping secrets away from the agent: what permissions can and cannot block
• 4:52 Which NRP model to use
• 6:43 Demo 1, the agent learns the project: /init, AGENTS.md, permission rules
• 13:48 Demo 2, asking and steering: a vague prompt against a precise one, plan mode, reading the diff, and data from Google Drive with gog
• 21:58 Demo 3, tests catch what you would miss: a failing test first, then the fix, in a pull request that closes the issue
• 29:54 Demo 4, continuous integration: GitHub Actions, the agent reads the CI logs with gh and fixes the failure, main protected
• 37:07 Demo 5, subagents keep the tutorial working: a tester and a fixer agent check the README in a loop
• 48:18 Recap and takeaways

The eight takeaways, as a checklist for any agent, are at the end of the setup post.

Questions and collaborations

Questions and feedback on the demos are welcome as issues in the demo repository. I lead the Scientific Computing Applications group at the San Diego Supercomputer Center: we help research groups build software for AI systems and high performance computing, and collaborate on proposals: see my SDSC page to get in touch.