Hello Reader
"Everyone keeps telling me to build agents and I don't even know where to start."
A marketing director said that to me last week and she wasn’t the only one.
We’re in full hype-cycle when it comes to AI agents. You can’t go anywhere without seeing that word splashed across digital displays, billboards and banners.
Even worse, your CEO read about them over the weekend and wants you to spin one up on Monday.
A vendor demo made it look simple but when you tried it for yourself it broke half way through.
Someone on LinkedIn says their agent runs the whole pipeline while they sleep, but doesn’t explain exactly how.
So you feel behind.
You're not behind.
Most marketing teams who jump straight to agents fail, and it has nothing to do with the agents themselves.
They fail because they didn’t build the foundation first.
This week I want to show you what "ready" looks like, using a real agent built by a real person.
Then I'll give you a prompt you can run today to find out where you stand.
Three Ways to Work With AI
Everyone wants to build AI agents, but ninety percent of the clients I consult, don’t even need an agent.
They need augmentation or automation.
Very few people are ready for full agentic work, and it helps to know the difference between all of these.
Here they are:
Augmentation is AI assisting you.
This is your traditional chat collaboration with an active back and forth. You tell Claude what you want, it gives you an output, you provide feedback and so on. You should use this model when you need room to explore, brainstorm or weigh options. This is a great model for drafting, summarizing or pressure-testing a plan before you take it to your boss.
Automation is AI running a defined job for you.
You give the AI explicit instructions and it follows repeatable, rules-based flows where the output is predictable. I have an automation I built in Make that triggers when a deal in Attio moves from working to won. It activates Claude, which then writes a customized contract based on the deal details and drops it directly in Google docs ready for me to add the eSignature fields. You use this model when you want to develop a predictable output every time.
Agency is AI running on its own.
You set goals, parameters, provide the context, and the AI makes its own decisions to reach the outcome. This might look like a daily briefing you get every morning in Slack or when Claude monitors your SEO keywords and delivers content recommendations to your project tracker. You use this model when you have a solid foundation — context, data, guardrails, success metrics and a very clear understanding of what a good outcome looks like.
Why Most Teams Aren't Ready
So what does "ready" require? Four things.
#1 Context and Data
An agent acting on your behalf needs to know your business the way a good hire does. Where your data lives, what's current, and what's clean. If your information is scattered across tools that don't talk to each other, the agent is working blind.
#2 Connected Platforms
An agent needs access to the tools where the work happens. As an example, if you don’t connect a project tracking agent to your airtable account, you’ll have no way for the agent to help you manage those projects.
#3 Instructions and Repeatable Setup
The agent needs to know what you want it to do. This requires a clearly defined workflow with detailed instructions for each step.
#4 Success Criteria and Guardrails
You need to know what a finished, correct result looks like, and you need a way to check the work. You also need to decide what the agent can do on its own and where a human signs off before it acts. Anything touching spend, customer data or publishing needs a checkpoint.
What It Looks Like When You're Ready
Adam New-Waterson leads the AI team at Orderly Wellness, a family of prescription medicine brands selling things like GLP-1s and HRT.
He came to AI sideways.
He was a CMO, moved into product, then taught himself to code with AI in 2023. He thinks like a marketer and builds like an engineer.
He told me that when he was hired, the company was on track to spend three million dollars a year on AI subscriptions with nothing tying it together.
His first move wasn't to build agents.
"I spent the first month just understanding what things would help the team be more efficient," he says.
First he built a workflow engine, and from there he started developing more advanced automations and agents. Now he has a complete custom workflow and agentic platform that helps the marketing team (and others in the company) answer market questions, generate campaign plans and create production-ready content including text and images, with FDA compliance checks built in.
It has a knowledge base of roughly 46,000 pages pulled from competitors, documents and research. Brand kits exist for every brand with logos, colors, typography and tone of voice. The team created custom personas with consistent AI-generated faces, so a marketer can place the same spokesperson in a new scene without hunting for stock photos.
Adam’s hardest problem was building an agentic advertising tool that would pull ICPs, brand context and auto generate a campaign across messaging, platforms, spend and creative.
No small feat to build.
He gave the problem to Claude Code and asked for an analysis. It told him what was working and outlined all the gaps he needed to fix.
Rather than doing it himself, he told Claude to tackle the gaps. "An hour later, it had coded the whole experience," he says.
What it built in that hour: a campaign strategist, fixes to the existing issues, three new pages and an embeddable script for attribution tracking.
Roughly six weeks of work.
To put a number on the week, Adam wrote around 240,000 lines of code, against a normal week of 15,000.
But Adam wouldn’t have been able to build this without the original foundation. For the ad agent to work it needs the knowledge docs, context and guardrails to produce an effective campaign.
What happens now? A marketer types a plain English request, "this is a $50,000 conversion campaign," and the tool creates the campaign, generates a media plan and pulls creative that's already on brand. It can do that because the brand knowledge, personas and data have already been loaded in the system.
Despite Adam working with Claude to build the agentic ad platform, the majority of the solutions he's creating aren't technically agents. They're either AI-assisted workflows (augmentation) or automations.
When one of Adam's product managers asked for an agent, he looked at the workflow and saw a simple yes-or-no decision.
"You don't need an agent," he told the product manager. "You just need a rule."
How to Find Out Where You Stand
Now that you’ve seen Adam’s agentic ad platform, you might be tempted to open Claude Code and start building your own.
I'd start with your foundation instead.
Once that's in place you can build a few simple automations, and when those run without breaking, it's time to tackle agents.
If you think you're ready, there's an easy way to test it.
Your AI already knows a lot about your setup. Which platforms are connected, projects and skills you have saved, all of its memory.
I built a prompt that puts that knowledge to work.
It runs in two passes.
First it scans what it can already see about your setup and grades its own confidence. Then it asks you the few things it can't see, like the task you have in mind and how you'd check its work.
At the end you get a verdict: ready, not ready or partway, and what to do next.