TECH 43
Supervising AI Coding Agents: Design, Test, and Trust the Output
AI coding agents can now generate applications from plain English, but their output is not always reliable.
They can make confident mistakes, fabricate code, rely on outdated dependencies, expose sensitive data, or build systems that appear to work until they fail.
This hands-on workshop introduces a practical supervise-and-verify workflow for using AI coding agents effectively and responsibly.
Students will examine how these systems work, where they tend to break down, and how clearer specifications, stronger prompting, testing, and human oversight can improve reliability.
Working with tools such as Google Colab, Claude Code, Replit, and OpenClaw, students will scope small projects, compare agent outputs, identify unreliable code, and assess when an agent’s work is trustworthy enough to use.
The workshop will also address API key hygiene, PII risks, and responsible AI practices.
Students will leave with a grounded understanding of the capabilities and limitations of AI coding agents, along with practical strategies for supervising, testing, and evaluating their output.
This course is designed for beginners and intermediate learners; professional software developers who ship code daily may find some portions of the material basic.
This course relies on the use of an external, third-party tool that is not managed or supported by Stanford.
Students must purchase their own tool subscriptions and can expect to spend $25–$100 per month.
Please see the course syllabus for more details.