Lesson 1 of 6 · Agents

What an agent is

The word gets used for everything from a chatbot to a robot. Here is the one-sentence definition, the four ingredients, and what is still missing even when all four are there.

In AI Basics lesson 3 you met the harness: the app wrapped around a model that gives it web search, files, and memory. An agent is what you get when the harness goes one step further and lets the model do things, on its own, across several steps, until a job is done. That is the whole idea. The rest of this lesson is about what each of those words hides.

The one-sentence definition

An agent is a model, plus tools, plus a goal, plus a loop. Four ingredients. Take any one away and you have something else: a chatbot, a plan on paper, a one-shot helper. Put all four together and you have software that can be handed an outcome and left to work toward it.

Model. The prediction machine from AI Basics. It reads text and writes text. On its own it cannot touch anything.
Tools. A tool is a function the model is allowed to ask for: search the web, read this file, send this email, run this program. The model does not run the tool. It writes a request, and the harness runs it. The next lesson is entirely about this.
Goal. An outcome rather than a question. "Why did our reviews drop this month?" is a goal. "Summarize this review" is a question.
Loop. The harness lets the model act, shows it what happened, and asks "what next?" over and over, until the model says it is done, or gets stuck, or hits a limit.

Tap the ingredients below in any order and watch what you get. Pay attention to the in-between states. They are the products you have probably already used without realizing they were almost agents.

Tap to add ingredients: chatbot to agent

You ask

Find out why our Google reviews dropped this month.

What you have

A chatbot

It reads your sentence and writes a plausible answer from memory. It cannot see your reviews, so it guesses: "Reviews often drop after a change in service quality or a busy season." Helpful sounding, and not about you.

Illustrative output. The reviews, the tech shortage, and the findings are made up to show the shape of the behavior, not a real result.

Agent versus chatbot

A chatbot answers. An agent works. That is the practical difference, and it changes what you can ask for. With a chatbot you break a job into questions and carry the answers from one to the next yourself. "Read this email. Now draft a reply. Now make it shorter." You are the loop. With an agent you describe the finish line and the loop runs inside the software.

Under the hood the difference is smaller than it looks. The model inside an agent is often the exact same model you chat with. What changed is the harness around it. It now has hands (tools), a finish line (goal), and permission to keep going (loop). The AI-for-you feeling of an agent comes almost entirely from that wrapper, which is why lesson 4 goes back to harnesses in detail.

One more thing a chatbot cannot do that an agent can: notice it was wrong and try again. Ask a chatbot for a spreadsheet formula and it gives you one. Ask an agent, and it can write the formula, run it on your actual file, see the error, and fix it before you ever look. That try-check-fix cycle is the loop doing its job, and it is where most of the value lives.

You have probably already used a small agent without the label. The "deep research" features in the big chat apps are agents with one tool (web search) and a fixed goal (write a report). Agent or browsing modes are agents with a browser for hands. The Agentic Processes module covers those as products. This module is about the general idea underneath them, which is what lets you judge the next one.

The new employee with a computer

The analogy that holds up best: an agent is a new employee who has been given a computer, a login, and a task. Think about what that actually means on day one at a Hurricane rental shop.

They are smart. They read fast. They can use whatever software you gave them access to. They will keep going on the task until it is done or until they hit something they cannot figure out. And they know nothing about your business except what is written down somewhere they can see. They do not know that the Johnson family always books the same two side-by-sides, or that you never rent on the Saturday of the Hurricane Enduro race, or that the "ask Mike" note on the whiteboard means the price is negotiable.

So the questions you would ask about a new hire are exactly the questions to ask about an agent:

What can they see? (Which files, which inbox, which systems.)
What can they do without asking? (Read, draft, send, pay.)
What have I actually written down for them? (Instructions.)
How will I know what they did? (Logs, a daily summary.)
What happens when they get stuck? (Do they stop and ask, or guess?)

You would never hand a first-day hire the company credit card and the keys and say "run the place." The same instinct applies here, and it is a good one. Lesson 5 turns that instinct into a checklist.

The hard part, simply

The model never touches anything. It only writes requests. This is the idea people trip on, and it matters because it is the reason agents can be made safe at all.

Think of a manager who is locked in an office with no phone and no computer, just a slot in the door. Slips of paper come in with information. Slips of paper go out with instructions: "pull the March invoices," "send this reply to the customer." An assistant outside the door reads each slip and decides whether to do it. The manager is brilliant, but every single action passes through that assistant. If the assistant is told "never send anything without showing the owner," the manager cannot send a thing. Not because the manager agreed, but because the manager physically cannot.

Worked example: you tell an agent "clean up my inbox." The model reads your inbox (a slip goes out: "read inbox," the assistant allows it, the emails come back on a slip). The model decides three emails are junk and writes: "delete these three." If deleting is allowed, it happens. If deleting is set to "ask first," the slip stops at the door and you see a prompt: "Delete these three?" The model's judgment was the same either way. What changed the outcome was the rule the assistant was given. That assistant is the harness, and those rules are permissions.

What the definition leaves out

Four ingredients make an agent. They do not make a colleague. Two things are missing, and both of them stay missing no matter how good the model gets.

Judgment about what matters. The model can be excellent at reasoning and still not know that the one-star review from a regular deserves a phone call, not a templated reply. It optimizes for the goal you wrote. If you wrote "reply to every review," it replies to every review. It does not know which one is your cousin. Everything it does not have written down, it does not have. This is not a flaw to be fixed by a smarter model; it is what it means to be given a task instead of living the job.

Accountability. When an agent sends the wrong quote to a customer in Washington, the customer does not call the agent. The signature on the email is yours. The license is yours. The refund is yours. No vendor's terms of service move that responsibility, and none of them try to. An agent can carry out a decision. It cannot own one. Keep that sentence in mind every time a product promises to "handle it for you."

Put together: an agent is a fast, tireless, literal worker with no context beyond what you gave it and no stake in the result. That is genuinely useful. It is also exactly the profile you would supervise closely, which is the whole subject of this module.

Try this yourself

You can feel the difference between a chatbot and an agent without installing anything. Open whichever AI app you already use and give it a goal instead of a question. Then give it a second goal that needs hands. Notice where it stops.

I own a small landscaping company in St. George. My goal: figure out which of my services to promote in September. Do not answer yet. First, tell me what information you would need to look at and what you would do with it, step by step. Then tell me which of those steps you can actually do from inside this chat and which ones you cannot.

The answer will usually be honest: it can reason about the plan, and it cannot open your booking calendar or last year's invoices. The steps it cannot do are the tools it is missing. Every one of those is a permission decision you would have to make before letting an agent loose on it. That list is your homework for the next lesson.

Next lesson2. Tools and permissions

Last updated August 24, 2026