One request, one reply, one confident guess — the model API request and reply, priced in tokens

A model API call is one request in and one reply out, priced and limited in tokens, and anything about your business that is not inside that request does not exist for the model — which does not stop it answering.

Scene 01

One request, one reply, one confident guess

Watch
Try it
Predict
Capture

Tinted lines are the request; the two real lines are what you wrote.

What to watch for

What actually happens when a program "asks an AI" something? Less than you would guess: it opens an ordinary connection to a server, sends one block of text, and gets one block of text back. Watch that call get assembled for Shopfront — the same online store as the metrics and tracing curricula, seen this time from its support desk. First the request appears: everything the program chose to send, including the standing instructions it puts in front of every customer message, and the customer's own line. Then the reply comes back. Then the same call is measured — providers bill and limit text in small chunks of about four characters, called tokens, so the bar beneath the stage is this call redrawn with length standing for quantity. Then the price lands top-right. Finally the same call runs again with a harder question, and you should watch what the reply does about a record nobody put in the request.

On the left: the request card — the entire block of text your program sends, with the lines that matter drawn as real words and the rest as tinted rules.
In the middle: the model, a lozenge with a stub at each end, because text is the only thing that goes in and the only thing that comes out.
On the right: the reply card — the text that came back, drawn the same way.
Between the request and the model: slate dashed boxes standing exactly where a lookup would have sat, marking a record the call never contained.
Under the stage: the same call redrawn as a bar, one segment per piece in the order it was sent — slate for the standing instructions, blue for the customer's message, emerald for the model's reply — on an axis marked in tokens — a token is the chunk of text a model is billed and limited by, roughly four characters of English prose.
The slate segment at the head of that bar is the system prompt: the standing instructions your program puts in front of every customer message, which the vendors spell system (Anthropic), instructions (OpenAI) and systemInstruction (Google).
Top-right: the cost-per-resolved-ticket readout — a coin stack for the money this call spent, and a tick or a cross for whether the ticket actually ended resolved.

Sources

Where this sits in Build a production AI agent (from one API call up)

Scene 01 of 2, in the One call act — Text in, text out. No memory. Check why it stopped.. A model API is a function from text to text: it knows only what is inside the request you send, it is billed by the token, and when it knows nothing it still answers fluently.

Up next. One call answered one question. The customer writes back — what does the model remember?

All 2 scenes in Build a production AI agent (from one API call up) · Every curriculum

Designs that use this

AI Agent PlatformLong-running multi-step LLM agents — durable workflow + sandboxed execution + LLM gateway. Brain / hands / state are independently replaceable. The agent run is a workflow, not a request.

Built with Arqly

Every scene in Build a production AI agent (from one API call up) builds on the one before it.All 2 Build a production AI agent (from one API call up) scenes