Human-Agent Engineering

Module 1, Lecture 1.2 | Introduction to Agentic Systems

Using AI coding agents as a concrete example, this lecture explores what effective human-agent collaboration actually looks like. We define "vibe coding" — giving an agent vague instructions and accepting whatever it produces — and examine why it fails for anything beyond throwaway prototypes. The alternative is human-agent engineering: providing clear context, reviewing output, iterating, and knowing when to take over. We introduce the principle that LLMs are only as good as the context you provide, and present the autonomy spectrum — a framework for deciding how much independence to give an agent based on reversibility, consequences, and capability.

Read the full lecture narrative

Additional Resources

• Lecture slides
• Andrej Karpathy's Original "Vibe Coding" Post — The February 2025 post that coined the term, describing a style of development where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists."
• Karpathy's 'Vibe Coding' Movement Considered Harmful — A detailed critique of vibe coding's failure modes, including loss of code comprehension, compounding technical debt, and the breakdown of engineering responsibility.
• Inventor of Vibe Coding Admits He Hand-Coded His New Project — When Karpathy released his "Nanochat" project, it turned out to be entirely hand-written — he found AI agents "just didn't work well enough" for his own real work.
• Vibe Coding - Wikipedia — A comprehensive overview of the practice, its origins, and documented problems including security vulnerabilities in production apps and studies showing AI-assisted developers were actually slower on familiar codebases.