EXHIBIT 01 / TOOL USE
How can AI operate software?
What actually does the work?
A model proposes a tool call by writing structured text. A program checks that request, runs allowed code, and returns the result. “The AI did it” hides the most important boundary.
THE SHORT ANSWERThe model writes. The program acts.
One action has four distinct parts.
- 01 / HUMANRequest
You describe an outcome in ordinary language.
- 02 / MODELProposal
The model writes a tool name and arguments. It has not acted yet.
- 03 / PROGRAMDecision
Software checks availability, parameters, policy, and permissions.
- 04 / TOOLEffect
Approved code changes state, then sends a result back.
Run the same request under two conditions.
First finish the default run with addTodo available. Then select “no tools,” finish that run, and compare the actual records. The model request stays the same; the program’s decision changes.
Make the program add one to-do
The request is fixed and nothing starts on its own. Advance one step at a time. Two of the six steps are computed by your browser as you go.
FIXED REQUEST Add “Read the AI primer” to my to-do list.
CURRENT STEP Step 1 of 6
InterfaceYou made a request; the list contains 0 items
Behind the scenesThe available-tool configuration is ready; nothing has happened
On a narrow screen, use these tabs to switch views. Switching a view never advances or reruns a step.
Add “Read the AI primer” to my to-do list.
The list is empty
- Tools available in this run
Available tools: addTodo
You Your request is present and the list is empty. Advance when you are ready; this page will not start by itself.
Up to this step, this page has not executed a tool
The actual record produced by each condition
-
The model is offered a tool that can write to the list Not run yet
Complete this condition to fill the card with its actual record. No result is filled in advance.
-
The model is offered no tools Not run yet
Complete this condition to fill the card with its actual record. No result is filled in advance.
SCRIPTED MODEL OUTPUT The tool request and both final messages were written in advance. No model is connected to this page.
LIVE BROWSER COMPUTATION Validation and addTodo execute in your browser when you advance those steps.
What was real, scripted, and omitted?
Model messages
The tool request, success reply, and refusal were written by the museum. No live model was called.
Validation and state change
Your browser evaluates the tool declaration and calls addTodo only after the request passes.
Production safeguards
This teaching system has one local tool. It omits authentication, authorization prompts, retries, concurrency, remote failures, and audit storage.
MECHANISM SOURCES
OpenAI function calling announcement source record Cloudflare Agents documentation source record The sources support the general request–execute–return mechanism. They do not claim that this museum’s scripted dialogue occurred in a real product.