Agent Foundations for Marketers · Lesson 1 of 8

What is an AI agent? A marketer's guide

Before any setup, the shift that makes agents click. An agent reads your files, runs your tools, and does the work, instead of just handing you text.

You have probably used ChatGPT or Claude for marketing work already. You ask for five subject lines and you get five subject lines back. Everything after that is still yours to do: pick one, paste it into the email tool, fix the details, send the thing. The chat only ever hands you text.

That is changing, because AI has learned to act. Give it a task and it can carry the task out: open the files, make the edits, save the result, show you what it did. The word for an AI working like this is an agent, and agents are already inside the products you know. ChatGPT can build documents, Claude can run code, Cowork works through folders you share.

So the difference that matters is not which product has an agent. Most of the big ones now do. The difference is where the agent works and what you let it touch.

AI agent: AI you delegate work to, not the kind you chat with. You brief it, it does the work: reading and writing your files, running programs, and, by default, checking with you before anything it cannot undo. The assembly gets handed off. The thinking stays yours.

Before we install anything or pick a tool, this first lesson builds the mental model. What an agent is, why where it runs matters more than which brand runs it, and what stays your job. Everything after makes more sense with this in place.

The developer-tool look puts people off

Here is the part that looks off-putting at first, so let me slow down on it. An agent can only act on what it can reach, and to reach your real files and programs it has to live on your machine. The tools built for exactly that already exist: they are the ones programmers use all day. So that is where the most capable agent setups live.

Some run inside a code editor, a program like VS Code or Cursor that shows your files down one side and a place to type on the other. Some run in a terminal, the plain text-only window that is all typed commands and no buttons.

The look is the intimidating part, and the look is all it is. You do not write code in there. You type instructions in plain English, the way you would brief a capable new hire, and read the answers back. The next two lessons walk you into one of these tools step by step, and it is far less of a wall than it looks from the outside.

What an agent can actually do

The basics you can guess: an agent reads and writes files, runs programs, reaches your other tools through connectors, and chains the steps on its own toward the outcome you set. A browser tab can do versions of all of that. The scale is what changes when the agent works on your machine, so here is what that has looked like in my work.

For one client we built full email campaigns and pushed them straight into OneSignal, the tool that sends them. The agent wrote the copy, built HTML that survives dark mode, and delivered everything through the API. Nothing was copy-pasted into the tool by hand.

The same setup built and now runs their full website, 800-plus pages. Every marketer knows the CMS fight: dragging blocks around Webflow, wrestling Framer, keeping an aging WordPress alive plugin by plugin. The agent replaced that fight with a folder of pages it edits directly.

Positioning tuned page by page, SEO and internal links maintained across the whole archive, imagery chosen and placed. A change that touches two hundred pages is one instruction, not an afternoon of clicking. The result is faster and more custom than the page builders allow, because nothing is limited to the blocks they ship with. A browser tab has no way to hold a system like that.

You give the agent an outcome, and it takes the steps in between. The size of the outcome you can hand over is the part that grows.

The environment decides what you get

The same agent behaves very differently depending on where it runs. In Cowork or the ChatGPT app, the agent works in the vendor’s workspace, on what you hand it: files you upload, the connectors it supports, results you download back out. For one-off tasks that is genuinely enough, and I still reach for those apps when that is all a job needs. But when the session ends, that workspace goes with it.

Run the agent in an editor, on a folder you control, and three things change.

You hand off bigger pieces of work. A boxed app can reach the tools it has connectors for. Real tasks cross tools: one flow I run pulls from Airtable, Google Drive, Directus, and a client’s own API, five sources feeding one piece of text. When one of those has no connector, just an API, the boxed app runs out of road. An agent on your machine writes the API call and keeps going.
Your standards apply themselves. Your playbook stops living in your head. A context file loads at the start of every session, and saved skills fire on their own when a task matches. You set the standard once, and it holds on your busiest day too.
Your memory is files, not an account setting. Chat apps have memory features now, and they prove the point: context is everything. But their memory is bound to your account. Yours is a folder, on your computer, on GitHub, wherever you keep it. Readable, editable, and yours to take to whichever model wins the next wave.

That gap is what the rest of this course is built on. Every lesson from here either makes the handoff bigger or the folder smarter.

Why this matters for marketing specifically

Most marketing tasks are ten minutes of thinking wrapped in an hour of assembly. You already know the article you want. Writing it means pulling the search data, checking what currently ranks, drafting, adding the internal links, and formatting the whole thing. The thinking is yours. The hour goes into the assembly around it.

Assembly is exactly what an agent is good at. It does not decide what to write or who it is for. What it removes is the distance between deciding something and shipping it. The reformatting, the data pulls, the first drafts, the jobs I used to save for a quiet afternoon, mostly happen in a few sentences now.

What an agent is not

A few honest caveats, because the gap between a slick demo and daily reality is real.

Not autonomous by default. The agent works from your instructions and, by default, asks before anything it cannot undo, like deleting a file or posting in public. Faster modes that skip those prompts exist, but you switch them on deliberately. The default keeps you approving as it goes, which is what you want from something that can touch your files and your accounts.
Not informed about your business. Out of the box the agent knows a lot about marketing in general and nothing about your positioning, your customer, or the constraints you have already learned the hard way. Handing it that context is real work, and it is most of what a later lesson covers.
Not everyone’s daily reality yet. The demos you see are the best thirty seconds of somebody’s week, edited. The daily reality is quieter, a lot of ordinary tasks getting done faster, with you checking the work. You are not behind, and the foundations here are enough to start compounding now.

Where this goes next

That is the whole mental model. An agent reads, writes, runs, and connects, all driven by plain instructions from you, working in steps toward an outcome you set.

From here it gets practical. Next we look at what an editor is and which tool fits you. Then we install one and open a first session on your own machine, so everything after that is you doing the work, not reading about it.