Agents, Not Chatbots

Most people who say they use AI at work mean this: they open a chat window, paste something in, read the answer, and copy it somewhere else. Forty times a day. The AI helped with the typing. The person is still doing the job.

This course is about the other thing. Systems where the work happens whether you are at your desk or not. By the end of this lesson you will have one mental model, and you will be able to point at any AI system and say which parts are plumbing and which part is the brain.

Chatbots answer. Agents act.

A chatbot is a conversation. You bring it a task, it responds, and the loop ends with you. Useful, but it scales exactly as far as your attention does.

An agent is a job description. It watches for work, decides what to do with each piece, and does it. The loop ends with a result, not a reply.

The difference is not the model. The same Claude that drafts your email in a chat window can run inside an agent. The difference is where the human sits. In a chat, you are between every input and every output. In an agent, you set the rules once and step out of the loop.

The three boxes

Every agent you will build in this course, and almost every agent you will ever see, is three boxes.

Trigger. Something happens. An email arrives. It turns 7am on Monday. A form gets submitted. A row lands in a spreadsheet. Triggers are dumb on purpose. They fire, that is all they do.

Agent. The model looks at what came in and makes a decision. What kind of email is this. Does this lead match our customer profile. Which numbers in this report changed enough to mention. This is the only box with judgment in it.

Action. Something concrete happens in the world. A draft is saved. A row is written. A Slack message goes out. Actions are dumb on purpose too. They execute exactly what the decision said, nothing more.

Here is the model on one line:

Trigger (something happens) -> Agent (AI decides) -> Action (code executes)

That is the whole diagram. When an agent misbehaves, the fault is in exactly one of the three boxes, and in practice it is almost always the middle one. Which is good news. You will learn to fix the middle box by editing text, not code.

A real one, taken apart

Here is an agent that runs on a real inbox at our company. Yesterday it processed 114 emails before anyone had coffee.

The trigger is a new email in Gmail. Every arrival fires the workflow once.

The agent step sends the email to Claude with one question, in effect: what is this and what should happen to it. The instructions name four categories. Urgent, needs a reply, newsletter, invoice. Claude reads and picks one.

The action step is a switch with four exits. Urgent posts to a Slack channel within a minute. Needs a reply gets a drafted response saved to Drafts, and a person reads it before it goes anywhere. Newsletters get archived. Invoices get labeled and forwarded to bookkeeping.

Of yesterday's 114 emails, 9 needed a human. The other 105 were handled, and nobody read them. That is the entire pitch for this course in two sentences.

The anti-pattern, for contrast

Now the same job, the way most people do it. Open inbox. Read an email. Copy it. Switch tabs. Paste into a chat window. Type "write a polite reply saying we will get back to them Thursday." Copy the answer. Switch tabs. Paste. Send. Repeat 40 times.

Count the human steps: about 8 per email. Call it 90 seconds each. That is an hour a day of being the glue between a mailbox and a model. The AI is doing the least interesting part, and the person is doing everything a trigger and an action would do for free.

If you recognize yourself here, good. Every build in this course starts from exactly this position.

The mistake, and the fix

Here is the part most courses edit out.

The first version of that triage agent used a one-line instruction: "Classify this email." Reasonable, right? Here is what came back on three real emails:

Email: "Invoice #4471 attached, due July 25"        -> "Business"
Email: "URGENT: server down, customers affected"    -> "Work email"
Email: "Our July newsletter: 5 automation tips"     -> "Informational"

Not one usable label. "Business" is not a category the switch knows. Nothing was wrong with the model. It answered the question it was asked, and the question was lazy.

The fix took two minutes. Name the allowed categories, say what happens to each, and give one example. The relevant part of the instruction became:

Classify into exactly one of: urgent, needs_reply, newsletter, invoice.
Respond with only the label.
Example: "Payment reminder: #2210 due Friday" -> invoice

Same three emails again: invoice, urgent, newsletter. Three for three, and the switch downstream knew what to do with each.

Hold onto that experience. When your agent gives you garbage in lesson 5, you will not think the model is broken. You will ask what you failed to tell it.

Recap

An agent is a trigger, a decision, and an action. The AI owns the decision and nothing else. If you are copying and pasting into a chat window, you are the trigger and the action, and that is the hour a day we are here to get back.

Next lesson: how to pick the right job to automate, and why the obvious candidate is usually the wrong one.