A trace is one piece of work, followed step by step through the stack. It shows what runs, in what order, what each step costs, where it runs, and whether a model, a system or a person does it.
The vocabulary around AI engineering arrives faster than anyone can absorb it, and a list of definitions is something people nod at and never read. This puts the terms to work instead: forty of them on one map, with real work moving through them.
Use it in this order
-
Run a trace
Pick one from the strip at the top of the page, then use the forward arrow to move through it a step at a time. Each step says what is happening, what it costs, and who is doing it. Press play if you would rather it advance on its own.
-
Follow a concept
Every term a trace touches is clickable. Each one gives you what it is, the products people actually reach for, what goes wrong when it is thin, and which other traces exercise it.
-
Use the reference below
The topology shapes, the vocabulary table and the nine diagnostic questions are there for looking something up rather than learning it.
What is real here
Model prices are per-token API list prices, and they are the only figures used for arithmetic, so every cost on screen is computed from token counts rather than asserted. They have not been checked against a live rate card, which makes them a basis for comparing one trace against another rather than a quote. A subscription is priced per seat rather than per token, so no figure here describes what a subscription costs. Check the vendor's current pricing before a number from this page reaches a client.
The scenarios are realistic but illustrative, and no client work is described. Products are named because naming a real tool teaches more than writing a vector database.
You can write your own trace as well. Describe a situation and Claude drafts the steps against these same forty concepts.