Why I'm Going All-In on AI Agents
I’ve spent the last few years going deep into AI and automations. Then in the last few weeks, everything changed. The models got smarter, the tools got real, and someone released an open source framework that brought it all together. I’m all in.
I’m building AI agents. Not chatbots. Not automations with a ChatGPT step bolted on. Actual agents that have an identity, memory, tools, and the ability to figure things out on their own.
What is an agent, actually
If you reduce it to the minimum, an agent is a stack of files on a computer. A folder with text and instructions. It has an identity (who it is), instructions (what it should do), memory (what happened before), and specific knowledge about its job.
Then you connect that stack of files to a brain. Could be Claude, could be GPT, could be Gemini. The brain is interchangeable. If one model gets better or cheaper, you swap it in. The agent doesn’t care.
On top of that, you give it tools. Web search, Shopify access, Meta Ads data, customer reviews, email, whatever makes sense for the job. Some agents have two tools. Some have twenty. It depends on what they need to do.
That’s it. A stack of files, a brain, and tools. The combination of smart models, big context windows, and all the infrastructure that developers have been building for the past two years creates something that genuinely figures things out.
Why now and not six months ago
This matters because six months ago, you couldn’t really do this. The pieces weren’t in place.
In 2022, we got ChatGPT. Input, output. Simple.
In 2023, the tools got better. Web browsing, bigger context windows. We went from “I tell you something, you tell me something back” to “I tell you something, you check some stuff, and then you tell me something useful.”
In 2024, we hit the automation era. Systems like Make and n8n that use AI at scale. A lot of people called these “agents” but they weren’t really. They were automations with AI steps.
Then at the end of 2025, the frontier models got a massive upgrade. Claude, GPT, Gemini, all of them, they got dramatically smarter at the same time. Nobody knows exactly why, but it completely changed the game. Because now these models also had all the tools and infrastructure that had been built along the way.
And then someone released OpenClaw. An open-source agent framework. One developer, gave everything away for free. It brought together everything that had been building up: smart models, tools, memory, identity, always-on capability.
That’s where we are. Right at the beginning of something real.
Where agents live
This is a practical question that matters more than you’d think.
You have three options. Your personal computer (good for you, useless for a team). A local machine at home (always on, but fragile, needs stable internet, goes down with the electricity). Or a virtual private server, a VPS, which is a piece of a computer living on the internet, hosted by a company.
For teams, VPS is the only reliable option right now. It’s always on, accessible from anywhere, has professional security and backup, and it scales. The cons are that it costs money (starting around 10 euros per month) and requires careful security setup. But there’s no serious alternative for running agents that a team needs to access.
How you talk to an agent
You talk to it like you’d talk to a colleague. Through Discord, Slack, Telegram, WhatsApp. You create channels for different topics. You structure it however makes sense.
I use Discord for my own agent. I have channels organized by project, by client, by topic. My agent, John, lives there. He has his own files, his own memory, his own tools. He knows who I am, what I’m working on, and what I’ve done before.
The way I think about it: imagine the smartest intern you’ve ever met. Always available, super fast, never forgets anything. If you had that person on Slack, what would you ask them to do? That’s what you ask an agent.
The types of agents I’m thinking about
There are roughly four categories, though the lines blur.
Autopilot agents. Small, trigger-based. If something happens on Shopify, the agent reacts. A QA agent checks a product page every day and flags problems. A competitor watch agent scans the market weekly. These run in the background. You set them up and they just work.
Thinking partners. Agents that know your business deeply and help you make better decisions. You ask “what are the best performing hooks for this product?” and the agent pulls Meta Ads data, cross-references with customer reviews, checks competitor ads, and comes back with an analysis. It doesn’t just retrieve data. It synthesizes.
Production engines. Agents that do substantial work. You give them a brief and they run through a framework, check multiple sources, and produce output. An SEO agent that analyzes Google Search Console, gathers competitive data, identifies opportunities, and then actually implements improvements on the website. A content agent that produces complete Shopify listings, product descriptions, and ad copy from a brief.
Platform agents. Agents that live behind a web interface. A tool that looks like normal software, but on the back end, every action sends information to an agent that uses tools and returns results. I’m building one of these for UGC video creation right now.
Read vs. write (this is critical)
When you connect tools to an agent, there’s a fundamental distinction. Read access means the agent can look at data but never touch anything. It can check a product on Shopify, read Meta Ads performance, pull customer reviews. But it can’t change anything.
Write access means the agent can create products, change prices, publish content, send emails.
This distinction is everything when it comes to risk. You start with read. You learn how the agent works, what it does well, what it gets confused by. Only after you genuinely trust it do you start giving write access, and even then, carefully.
The risks are real
I’m not going to pretend this is all upside. There are real risks.
Prompt injection. Someone tricks the agent into breaking its rules. An agent reads something on the internet that contains hidden instructions, and it acts on them. This is why guardrails matter, why you tell the agent explicitly what it should never do, and why you start slow.
Security. Your agent lives on a server. That server can be attacked. Big companies like Shopify have thousands of engineers securing their systems. Your VPS has whatever security you set up. I’ve done a lot of research and testing on hardening, but at some point, if you’re handling sensitive data, getting a security engineer to review everything is worth it.
Agent mistakes at scale. An agent doing something wrong once is annoying. An agent doing something wrong a thousand times before anyone notices is a disaster. This is why human approval before anything goes live is non-negotiable at the start.
Technology shifts. Something easier, cheaper, or better could emerge tomorrow. What we’re building now could become obsolete. My bet is that the knowledge, the frameworks, the SOPs, all the thinking we put into structuring how agents work, that transfers. The specific tool might change. The understanding doesn’t.
Why I’m going all in
Because the intersection of “understands marketing deeply” and “can build agent systems” is a very small space. Most people building agents are engineers who don’t know what the agent should actually do. Most marketers are intimidated by the technical side.
I spent years writing copy, running creative tests, doing customer research, building funnels. I know what good marketing work looks like because I’ve done it. Now when I build an agent, I’m not just connecting APIs. I’m building something that does the work the way it should be done.
I don’t know exactly where this goes. Nobody does. The technology is weeks old in terms of real capability. But I think the people who are in the middle of it right now, actually building and deploying agents, not just talking about them, will understand the future first.
That’s the bet I’m making. I’ll keep sharing what I learn along the way.