This is an explainer, not a build-along. No exercises, no labs, no code to copy: you will not build an agent here. The last module opens with the argument the field is having right now about calibration, and that lesson alone is worth the price. What you get instead of a build is understanding. You will understand the decision layer inside an AI agent well enough to read someone else’s, argue about it in a design review, and know what it costs per hop. If you want a hands-on build course, buy one of those instead, and this one after. Saying that here, before you buy, is the point: the field sells hands, and this course sells the map that makes the hands worth anything.
Most courses about AI agents route every decision through a large model and never ask what that costs. This one takes the opposite route. A System One model is a fast, typed decision layer: you hand it a state, you ask it questions, and typed answers come back. No parsing, no invented fields, no prose to unpick. Jev is the implementation used as the worked example throughout. What makes this course different. Four things, and each one is checkable against the curriculum.
Example in full. Every module follows one case from beginning to end. Modules four and five follow a single agent through its rebuild and then through every way that rebuild went wrong, which is more than a third of the course spent on one story. Honesty, structurally. There is a whole module on failure: confident wrong answers, a counting question returning the wrong number, dates read as literal text, a yes/no and a choice contradicting each other on the same input, and injected text moving a decision. The vendor’s own limits page is the source.
