The model asks. Your code is the one that acts — the tool call, the gap, and who executes it

It does not: the model emits a structured request naming a tool and its arguments and then stops, and every consequence — the lookup, the money, the record of it — happens in your code afterwards, in the gap between the model's output and the tool.

Previously

A schema fixes the shape of a reply, never the truth of the values inside it — but a shape your code can parse is still worth having. Now that object arrives asking for a refund. Who actually moves the money?

Scene 05

The model asks. Your code is the one that acts

Watch
Try it
Predict
Capture

The wall divides your model side from your tool side.

The chip touches the wall. Nothing has passed through it.

What to watch for

How does a model that only writes text look up an order or move money? It does not — not directly, and that is the whole scene. The order-status ticket from the support desk is back: "where is order #48213?" Watch two beats. First, what comes out of the model is not an answer: it is an amber chip naming one tool and its arguments, and it stops dead at the wall running down the middle of your program. Then your code passes it through the notch, the line reaches lookup_order, and what the tool returned comes back across the wall and lands on the window bar as a new message. Watch which side of that wall each beat happens on.

Across the middle: a rounded frame labelled your code — everything inside it, the model on the left and the four tools on the right, is your own program.
Down the centre of that frame: the wall, broken only by a notch where a line is allowed through, so every line between the model and a tool crosses there and nowhere else.
On the wall, when your code checks the call before running it: a small block the line has to pass through.
Left of the wall: the model, and pinned under it the card holding its actual output as real text.
The amber chip leaving the model is a tool call — the structured request the model emits naming one tool and its arguments, after which the response ends and nothing has run (vendors spell it tool_use with an input object, or function_call with an arguments string your code has to parse).
Right of the wall: the four tools in fixed order — lookup_order and search_policy outlined because they only read, issue_refund filled with a lock because it moves money, send_reply filled because it leaves the building.
The rose line coming back is a tool result — the message your code appends carrying what the tool returned, tagged with the same id the call carried, which is how the two are matched (tool_result with tool_use_id, or function_call_output with call_id).
One exception worth naming once: provider-hosted tools such as web search run on the provider's own machines inside the same response, so for those there is no gap for your code to act in.
Along the bottom: the window bar, one segment per message in order, one unit per token — and the violet segment near its head is the four tools' descriptions, which ride along in the same request (300 tokens here, an example figure).
Top-right: the example cost of this ticket — coins over a tick when it ended resolved, over a cross when it did not.

Sources

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

Scene 05 of 2, in the Tools & loop act — The model asks, your code acts, and then it loops.. A tool call is a message the model emits before it stops; your code decides whether to run it, runs it, and appends the result — so until your code acts, nothing has happened anywhere.

Up next. One request, one tool run, one result. Who decides whether there is a next step?

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