Marketer’s Guide to Agents: Why Your AI Work Is Still a Back-and-Forth Conversation

What AI agents are, how to split marketing work across them, and how to set up your first one.

This may look familiar if you use AI at work every day. You need a new product page, so you open a chat and ask for an outline. You read it, ask for changes, and then ask for the copy. The copy comes back sounding generic, so you paste in your brand guidelines and go another round. Then you move the result into your website editor block by block and go back to the chat for the launch email. The AI did real work in every turn, but you carried the context between turns, and most of your time went into moving the work along.

AI agents are how you move past working turn by turn, because they have AI do the work for you and complete longer, multi-step tasks. Claude, ChatGPT, and Copilot all offer a version of this. Two things about agents aren’t obvious, and they keep people in the back-and-forth. The first is that an agent is a text file of instructions that any chatbot can create for you, so creating one doesn’t necessarily take technical skills. The second is that the interface is still a chat box, so nothing tells you that you should ask for one.

What is an AI agent?

An agent is a set of instructions for one job, plus the tools it’s allowed to use and the things it’s never allowed to do. A chat waits for your next message, while an agent takes the job, works through it without back-and-forth, and reports back when it’s done.

A copywriting agent is a good example. Its file tells it to write from your product’s positioning, points it to your brand voice and writing rules, and says it’s never allowed to invent a claim or a number. You ask it for landing page copy, and it writes the page, checks every claim against its source, and hands it back with a list of any claims it couldn’t back up.

How to split marketing work into agents

Think about creating agents the way you’d think about hiring people to join your team. You break the work into similar chunks and give each chunk to one person. If you’re running a content marketing team, you’d give someone the job of writing copy, someone else the job of creating custom graphics, and someone else the job of publishing or project managing the team. Each of those would make a good agent on its own.

The reason to split these up into different roles is because we give agents each their own set of instructions, tools, and permissions. A copywriting agent needs your brand voice and has no reason to touch the website, and a web agent needs the website but doesn’t need your keyword data. Scoping each agent this way keeps it focused on one job and unable to change anything outside of it.

This is how the work splits in my setup, which runs in Claude Code. A skill is a saved set of instructions for one task.

• Copy: Writes marketing copy from the product’s positioning, using the brand voice skill, writing rules, and layout skill. It never invents a claim or a number.
• Web: Does the publishing work, building pages on the staging site, the private copy of our website, and QAing them at desktop and phone widths. It uses WordPress command-line tools, a browser tool for screenshots, Recraft for images, and layout and design system skills. It never clicks the final publish button, which is a rule I set.
• SEO: Picks the focus keyword, optimizes headlines, writes the SEO title and meta description, and adds internal links, using DataForSEO and Google Search Console. It never changes the address of a published page.
• Marketing automation: Builds programs, emails, lists, forms, and campaigns up to the send, using Marketo and a written playbook. It never sends to contacts.
• Product marketing: Maintains the positioning for each product, using research tools and a positioning toolkit skill. It never sends questions or signs off on decisions.

What a coordinator agent does

A coordinator is the agent you talk to, and it manages the others. It handles every request the same way.

1. It takes your request and your inputs, like a sales deck and your notes from a call.
2. It breaks the request into parts and puts them in order, such as positioning, copy, layout, build, and SEO.
3. It writes a complete brief for each agent. Agents can’t stop to ask questions partway through a job, so anything the coordinator can’t fill in comes back to you as a gap instead of a guess.
4. It starts each agent, and each one reports back with what it did, what it assumed, and what it couldn’t check.
5. It checks each report against the brief and sends anything missing back to the same agent.
6. It brings you the finished work, along with the decisions that are yours to make.

What AI agents need to know

An agent can produce larger chunks of work or even full tasks, and it does that by using the context in the folders and files you give it access to. That’s a different scope than what you’d typically upload to a web chat, and it’s certainly not your whole computer. Agents, including your coordinator, can pull what they need from those folders at the moment they need it.

Consider a copy agent working on the landing page for an event. If it needs specific information while writing, like the venue or dates, it can search through the other files in those folders, including ones it created itself, to find what it needs instead of giving you a placeholder and asking you to come back to it later.

For marketing work, three kinds of files are worth setting up first.

• Positioning: My toolkit for a product is about 20 files on who buys it, the problems it solves, how it’s positioned, and what we can prove, with every line traced to its source. New material gets folded in as it arrives.
• Brand and writing rules: Once these are written down, every correction you’ve made about how your team writes gets applied without you repeating it.
• How your website is built: A skill that documents your site’s layout blocks lets the copy map to the page before the build starts.

Building a product page from a sales deck with AI agents

I needed a page for a new product feature. I was provided a sales deck that outlined some of the key features, along with context from a quick recorded call with my CMO. I gave the coordinator both and asked for a first version on staging. This is what happened from there.

1. The product marketing agent folded the deck and the call into the product’s positioning toolkit. It marked the claims with nothing behind them, and it put open questions on a list for the product owner to get specifics on, which would be used for future iterations.
2. The copy agent wrote the copy from the toolkit, following our brand voice and writing rules, and left the unbacked claims off the page.
3. The copy agent used the layout skill to map the copy to the page’s blocks in a build sheet, so the copy was ready to fit the templates on the site and we didn’t end up with copy that doesn’t work for our layouts.
4. The web agent built the page from the build sheet as a draft on staging and QAed it at desktop and phone widths, using screenshots to check its own work.
5. The SEO agent set the focus keyword, SEO title, and meta description.
6. The coordinator checked each agent’s report against its brief and brought back the page with a preview link and a short list of what had been assumed and what couldn’t be verified.
7. I made the design calls, chose where to link to the page from, and clicked publish on staging myself, which is a rule I set rather than a limit of the tools.

How to set up your first AI agent

1. Write down one job you do all the time, with what good looks like and the steps it takes.
2. Ask your AI tool to set up an agent for that job, and have it grill you before it builds anything. You could say “I want an agent that builds pages on our staging site, and before you set it up, ask me about every decision it’s going to need to make.” The questions it asks become the agent’s instructions, and it’s the same approach I covered in Why Your AI Output Sounds Generic, and How to Fix It.
3. Decide which tools the agent gets and what it’s never allowed to do.
4. Put what it needs to know in a folder, and give the agent access to it.
5. Run it on real work, and when it misses, add the fix to its instructions so the next run starts with it.
6. Repeat with the next job. Once you have a few agents, ask for a coordinator with subagents it can hand work to, so each request starts from what’s already built instead of a blank chat.

Summary

A chat leaves you moving the work from one turn to the next, while an agent takes a whole job and reports back when it’s done. Treat each agent like a role you’d hire for, with its own instructions, tools, and permissions, and give it access to the folders and files it needs so it can find information without coming back to you. A coordinator runs those agents in order and brings you the finished work, along with the decisions that are yours to make. Start with one job you repeat, and add the next once that one holds up.