0b · From Chat To System — AI Systems Orchestration

Giving The Model Hands

The last chapter ended with a model that can only pass notes through a mail slot. So how does an AI search the web or book a meeting or edit a file?

It doesn't. Your code does. The model asks.

This is called a tool call. It works like a handshake. You tell the model in plain text which actions are available. Here is a search tool. Here is a send-email tool. Here is what each one needs. When the model decides it wants one, it does the only thing it can do. It writes text. But it writes a very specific shape of text. A small form: run the search tool with the query "flights to Denver". Your code reads that form and runs the real search itself. The results go back to the model as more text. The model reads them and keeps going.

The model never touches anything. It asks. Your code acts. The results come back as words. Every AI product that appears to do things in the world is this handshake repeated.

Notice what that means: your code decides which tools exist and what they are allowed to touch. That single sentence is the seed of every safety conversation in this book.

FIG 0b.2The tool-call handshake. The model asks in text. Your code acts in the world. Results come back as text.

The Word Agent, Finally

You have now heard the word agent used forty different ways. Here is the one honest sentence underneath all of them.

agent=a model call in a loop with tools until done

A chat is one lap. You ask. It answers. Done. An agent keeps lapping. It looks at the goal and takes an action with a tool. It looks at what happened and decides what to do next. It keeps going until the goal is met or something stops it. That is the entire difference between chatting and an agent. The loop.

One more distinction and you can follow almost any conversation in this space. In a workflow, you decide the steps in advance and the model fills them in. In an agent, the model decides the next step as it goes. Workflows are predictable. Agents are flexible. Real systems mix both. Choosing the mix is a running theme of this book. Chapter 04 expands the distinction. The two sentences you just read are enough for now.

Check The Work

Here is the habit that separates a demo from a system. It is almost embarrassingly simple.

After the model acts, something checks.

The check can be tiny. Did the code it wrote actually run? Did the answer come back in the exact shape we asked for? Does a second cheaper model call agree with the first one? Did a human glance at the email before it went out?

An example makes it concrete. Suppose you ask a model whether a support ticket is urgent. It replies "Yes, this looks fairly urgent to me." Friendly. Useless. Software cannot safely act on a vibe. Ask it instead to answer in a fixed shape and it returns {"urgent": true}. A program can check that answer and log it and act on it. Same model. Same question. The difference is that one of them can be verified.

Demos skip the check and look magical for one afternoon. Systems keep the check and stay useful after you stop watching them. Later chapters talk about evals and guardrails and approval gates. Every one of those is this habit with more machinery behind it.

Your First Shape

You now know enough to read the smallest useful AI system. The book returns to this shape over and over:

take the task in
  -> gather what the model needs to see
  -> let the model act
  -> check the work
  -> keep the evidence

Take the task in. Gather the right context because of the goldfish rule. Let the model write or act through its tools. Check what came back. Keep a record so a human can trust it later.

Five steps. No framework required. Most of the impressive systems you will meet in this book are this shape with more care at each step.

Later chapters compress those five steps to intake, context, action, verify, artifact. Same shape. Shorter words.

The Full Loop

A real system is the same loop with more parts:

A goal states the work to be done.
A planner decides the next lap.
A context builder gathers what the model needs to see. It handles the goldfish rule.
A model router picks which model runs the lap.
Tools take the action.
Artifacts are the record that it happened.
Evaluators check the work. A human approves the steps that have consequences.
Memory carries lessons into the next run.

That is the whole book: a loop that acts and checks its own work and leaves evidence.