Everyone is obsessed with autonomous agents right now, but you don't always need one. Often, the better approach is to use AI to generate structured data up front, then let standard code handle the rest. Keep it simple: use AI to build the setup, and run simple logic at runtime.

To prove it, I built an AI movement coach that breaks down videos (like a tennis serve, dance move, or martial arts kata) into step-by-step interactive lessons. A live webcam coach then tracks your form and scores you as you practice.

I built the app using AI tools, but at runtime, it only makes a single call to Claude per new lesson (and even that can be skipped). Once a lesson is created, the system caches it and converts each movement into target joint angles. Plain math handles the scoring in real time. The result is fast, cheap, predictable, and simple.

In this talk, I'll walk through:

Live demos of the coach in action My AI-assisted development workflow Practical strategies to keep runtime costs low Honest takeaways on when an agent is actually worth it versus when a single API call does the job