Loading
You have already built the easy version of this. HouseWarm shows a per-field confidence number, the source crop and the raw OCR text, and a human approves before anything downstream fires. That is one value, from one model call, in one gate. The thing the next job asks for is a different problem entirely: a run that took eleven steps, called four tools, retried one of them, cost sixty cents, cited three documents and changed two records — rendered so that a reviewer can tell in four seconds whether to trust it, and in four minutes exactly where it went wrong.
Start lesson one: The view comes second →
Almost every team builds that panel by drawing it first. This course does the opposite, and the ordering is the whole argument: learn the data shape before designing the view. A trace renderer designed against an imagined payload is a renderer that handles the happy path and quietly loses the retry, the parallel call and the run that died at step seven — which are the only runs anyone opens the panel to look at.
The whole four-module spine is registered so the lesson numbers never move. The first module — the data layer everything else renders — is written; ask your teaching agent for the next one once <AgentTrace /> has a schema it can actually parse.
What the runtime actually emits, learned before a single pixel of the trace panel gets designed.
The four altitudes as one component family, from a one-line summary to the raw payload.
Tying a claim to its source, and saying how sure the system is without inventing precision.
The numbers a reviewer needs per step, and the roll-up the person paying for the run needs instead.
Every claim on these pages links to its source. If a source looks wrong or out of date, check the resource list and tell your teaching agent — the course is meant to be corrected.