Your first AI agent
The same three parts as the last lesson, except the middle one decides what to do. That single difference is what separates a rule from an agent.
The last automation was a rule. A deal moved, so an email went out — and it would have gone out in exactly the same shape every time, because nothing in it was deciding anything.
This one is bolder. A list of leads nobody has called; an AI that rings them, has a real conversation, works out whether they are worth pursuing, and writes what it learned back to the CRM. Same three parts — something starts it, something happens, something is different afterwards — with a very different middle.
No workflow on this page. A voice agent needs a telephony provider connected, which the tutorial organisation does not have — so rather than show you a workflow nobody ran, this lesson describes the shape. The course has two agents you can see in full later, both with real workflows: a WhatsApp support agent, and the Core Agent lesson it is built on.
What is different about an agent
| A rule | An agent |
|---|---|
- An agent
- An agent
- An agent
- An agent
The last row of that table is the one people underestimate, and the last-but-one is the one that decides whether a project succeeds.
The shape
- Something starts it. A schedule, or a list of leads that meet a condition. Same kind of trigger as the last lesson.
- An agent takes over. You brief it once — who it is, what it is trying to achieve, what it must not do — and it holds the conversation.
- It has tools. Things it can do rather than say: look a contact up, write to the CRM, search your documents, hand the call to a person. It chooses when to use each one.
- Something is different afterwards. A qualified lead, a note on the record, a call transferred to somebody who could help.
The important word is “chooses”. You do not draw a line from the agent to “look up the contact”. You attach the tool and describe when it is useful, and the agent decides. That is genuinely powerful and it is also the reason agents need more care than rules.
Two things to hold on to before you build one
A prompt is not a control. If a rule must hold — never discount past this figure, never promise a date, always escalate after two failed attempts — it belongs in a decision node that the workflow enforces, not in a sentence you hope the model follows. This is the single most useful idea in the whole AI half of this course, and it is easier to accept now than after something goes out.
An agent that is not bounded will not stop. How many steps it may take, how long a call may last, how much it may read — every one of those has a limit and every limit has a default. The defaults are usually generous. Knowing which one actually ends a runaway conversation is worth more than any amount of prompt tuning.
What this course does with that
You will not build an agent for a while, and that is deliberate. Agents are made of the ordinary nodes — a lookup, a search, a write, a message — and an agent whose tools you do not understand is not debuggable.
So the order is: the Builder, then triggers, then the ordinary actions, then logic, then agents, and then two complete agents assembled from all of it. By the time you reach the WhatsApp agent, every node in it will be one you have configured yourself.
Next: The canvas — the screen all of this is built on, and the six things on it worth knowing by name.
Related lessons
Base rates — what a piece of evidence is actually worth
A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.
ReadConfirmation and survivorship — what you never looked for
Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.
ReadLoss aversion, sunk cost and regression — what it costs you
Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.
Read