Over the past four months, I have spent more than 700 hours building AI agents for my businesses. Not experimenting with AI — building production systems with it. This is the first of a four-part series on what that process actually looked like: what broke, what worked, and what I would say to another operator considering the same path.
A quick note on context. I run Marching Dogs, a media and commerce platform covering the resale industry, and I’m a longtime operator in retail arbitrage through House of Carts. Marching Dogs has always been about pulling back the curtain — explaining what is actually happening under the hype. In our Technology section, that now includes documenting how a non-engineer is building real AI infrastructure for a small business, and what that experience reveals about where this industry is heading.
The Problem That Started It
I did not come to AI because of a whitepaper. I came because I had too much work.
Anyone running a small operation understands the pattern. Weekly release setups. Inventory reconciliation. Spreadsheet formatting. Customer data cleanup. These are tasks that should not take hours, and yet most weeks, sixty percent of my time went to formatting instead of thinking. The question was never how do I use AI. It was what can I stop doing by hand.
For nearly two years, I paid twenty dollars a month for ChatGPT. It was useful for brainstorming, for editing thousand-row spreadsheets, for parsing data when I framed the question precisely. But the outputs were not always trustworthy, and the token budget at my price tier capped what I could deploy at any real scale.
In mid-2025, after persistent encouragement from a developer friend, I tried Claude — specifically, Claude Code. My first serious build was Pawket Change, a budgeting, inventory, and wealth-tracking application for small business owners. My wife and I had spent years cycling through budgeting apps and never found one that fit. Neither of us had the discipline to maintain the process in a spreadsheet. So I built what we needed.
It took two weeks, working 8 PM to 4 AM, driven by curiosity more than caffeine. I have never written code professionally. Traditional software development had always felt inaccessible — pattern-matching syntax, debugging across files, carrying the whole system in my head. Claude Code changed that calculus. I could describe what I wanted, plan carefully, and watch the product take shape.
A note on accessibility. Opening the door for builders from non-technical backgrounds is something we are actively exploring at Marching Dogs. If you have ever been told you “can’t code,” this series is worth following.
From Tool to Teammate
Pawket Change was the proof of concept. Not a demo — a production application, more than 100,000 lines of code, running daily. For someone without a formal engineering background, that was the inflection point where AI stopped being a chatbot and started becoming leverage.
By the end of 2025, I had eight application concepts in planning and three in active development. The first quarter of 2026 is when I stopped using AI and started building with it. I deployed a multi-agent framework across several experiments: agents responsible for planning, research, content, code review, and operational work that could continue when I was not at my desk.
That shift matters for a specific reason. As an operator juggling multiple applications, a resale community, Marching Dogs, a family, and a day job, I do not have more hours to give. Agents that can execute on my behalf are not a productivity hack. They are the difference between shipping and not shipping.
Where It Went Sideways
I will be direct: I have had multiple failed deployments. Not one — several.
The pattern was consistent. I moved too fast. As a solo operator, I can absorb risk that larger teams cannot, and I took the risk aggressively. In most cases, the failures were less about strategy and more about curiosity pushing me past what the underlying software was ready to do.
My first agent framework was called OpenClaw. I ran more than two hundred agents on it, each with its own role and capability. At one point Claude Code observed that what I had built was less a company than a small government. The observation was fair. OpenClaw was a prototype framework that was never designed for that scale. I was asking a concept car to run the Daytona 500.
The most expensive mistake was not financial. It was time — 700+ hours across four months, much of it spent pushing software past its present limits and learning, slowly, where the real ceiling was.
I also consumed tokens at a rate I would now call reckless. Most users consider 2 to 5 million tokens per month heavy usage. I was running approximately 780 million per month and approaching one billion before I forced a step back to find cost-effective solutions that would not bury me.
When the second OpenClaw deployment was winding down, I planned a third — and then scrapped that plan in favor of Hermes Agents, a more secure and robust framework built by NousResearch. That single infrastructure change resolved most of the issues I had been fighting. Sometimes the right move is not a better plan. It is a better foundation.
What I Would Tell Another Operator
If a peer asked where to start, my first answer would be this: do not use AI for reasoning when a script will do.
Burning LLM tokens on repetitive, systematic tasks is one of the largest cost drivers in the space. Operators pour money into reasoning-heavy calls for work that a twenty-line script could handle. Identify the repetitive work first, then ask a harder question: does this actually need a large language model, or does it need conventional automation that an LLM helped you design?
Depending on the nature of your data, there is also a credible case for running or training models locally. Private, in-house, predictable costs — nothing billed by the token, only electricity and time.
The Real Lesson
What this experience actually changed is how I value my time.
A few focused hours with AI can now produce what a team of engineers would once have needed months to deliver. All of it without a computer science education. That reframes what is worth doing, what is worth delegating, and what is worth purchasing.
The most meaningful outcome has not been the applications I have built. It has been what I can now give back to the communities I operate in. At House of Carts, a custom agent now handles our internal knowledge base — our guides, our partner tools, our member discounts. Members get accurate answers in seconds. Our team gets their evenings back. The operation runs better.
Before any of this, I was burning out. Ten years in retail arbitrage. Consulted by the New York Times, the Wall Street Journal, Vibe Magazine and others on where this industry was going. And yet I was running out of next moves. AI had felt like a pipe dream — I had assumed my inability to code was a permanent ceiling.
It was not. And the hunger came back.
The Anthropic Moment
Anthropic recently announced limits on OAuth token connections for third-party harnesses. As a heavy user, I understand the move. Their servers are not infinite, and premium pricing for outsized loads is a reasonable response. I was likely a small contributing reason; if every user ran a billion tokens per month, no provider could serve them.
But the signal is unmistakable. Hosted AI at scale is going to become more expensive. For most small businesses, the path forward involves investing in hardware that can run AI locally and privately. Own the infrastructure. Control the cost. Keep the data. It is part of what I am already building toward.
An Honest ROI Assessment
Two months ago, I would have said the time had not paid off. Today I would say the opposite. The current system has durable memory, continuity across machines, and the ability to continue working on my behalf while I am away from my desk. That is a return you cannot purchase as a subscription.
In earlier builds, I would get three or four days — sometimes a week — of clean output before the system began to collapse. Since moving to Hermes Agents, I have been able to grow the system without breaking it. Stability is what has made the return on investment real.
Will AI Replace Employees?
No. Not for small businesses. Possibly in narrow ways for some larger organizations. But not broadly.
The real future is equipping humans with AI to substantially multiply their output, not replacing them. Any company that believes it can eliminate a full department and substitute an AI is about to learn a costly lesson.
Without people who genuinely understand the business, AI is useless. When the model drifts or hallucinates, nobody without domain expertise will catch it. We have already seen the consequences. Duolingo replaced human instructors with AI and watched user engagement, time-on-app, and brand reputation all decline within a month. They are not the only example. They are the loudest one so far.
One Year From Now
In April 2027, what I want to be able to say is this. I built things I never thought were possible. My persistence pushed the people around me — including senior engineers — to ask what they could do. And I gave the communities I operate in, House of Carts and Marching Dogs, the tools to do the same for themselves.
Because the real win is not the technology. It is the moment someone without a coding background ships a real application, and the person next to them thinks: maybe I could too.
Five Quick Lessons
- Best AI tool I have used: Claude Code.
- Worst money I spent on AI: Udio AI music trial. Runner-up, an annual Cursor subscription — this space moves too fast to prepay a year on anything.
- Something AI still cannot do that surprises me: Honestly, very little. When I cannot get something done, it is almost always because I did not know how to ask.
- The project I am most proud of: My agent team.
- Advice in ten words: The best time to start was yesterday. Today works.
Continuing in this series
- Part 2: Things I Did Wrong Building AI Agents (So You Don’t Have To)
- Part 3: A Smarter Path — What I’d Do Differently Starting AI from Scratch
- Part 4: Your AI’s Memory Will Outlive Your Tools
If you are currently building with AI, or weighing whether you should be, I would be glad to hear from you. This series is written for you.