I Sat on an Idea for 7 Years. AI Helped Me File for a Patent in 2 Weeks.

Seven years ago, I was sitting at a bar after work, having a quick beer to relax before heading home for my second shift with my newborn son. Or maybe at that point he wasn't so newborn, but he was still a baby. An older lady sat next to me with a dog. The dog was very docile, probably an older dog, just lying on the floor, waiting for its master to finish her drink. When the lady got up to walk the dog out, I saw how the leash had become tangled around the dog, and she spent a minute untangling the mess. Watching that, I came up with the idea for a better dog harness. For the next few months, maybe a year, I worked on the idea on and off, but then it died because the next hurdle (actually creating a prototype and filing a patent) was just too much to get through after dealing with all the demands of day‑to‑day life.

But I didn't let it go. Not entirely.

How I started, and then stopped

I sketched it out. The fix in my head was simple: don't fasten the leash to a fixed ring on the harness. Put a cable along the top of the harness, and let the attachment point, a small carriage, slide along that cable. Dog crosses behind you, the carriage slides with it. Nothing to knot, because the anchor point moves instead of the dog wrapping around a fixed one.

FIG. 1 from the provisional patent: the sliding carriage, front view. This is the actual 3D-printed prototype. The leash clips to the ring on top, the ring in the middle rides along the cable that runs the length of the harness, and the lips and bottom wing keep it captured in its housing.

FIG. 2: the same carriage in perspective. The loop on top is where the leash attaches; the channel underneath threads onto the harness cable.

I spent a few months on it. I designed the carriage, and I 3D-printed it. I kept iterating on the design and bought the material for a simple prototype. I even bought a large stuffed animal dog since I was not a dog owner, so that I could fit a harness on it for testing. It was big and realistic-looking enough that after it first arrived, I simulated it barking and I scared the shit out of my young daughter. Her memory of that moment is not as funny as mine is.

After about a year of on-and-off working on it, I stopped. Not because it didn't work. Because the next steps weren't the fun part. Prior art searches. Figuring out whether it was even novel. Technical drawings that a manufacturer would take seriously. And the one that actually killed it every time I picked it back up: filing a provisional patent. I'd read a few guides, get intimidated by the specification requirements and the legalese, decide I'd "do it properly when I had time," and shelve the whole thing.

I did that for seven years. The printed carriage sat in a cardboard box with other parts and pieces of a prototype in a cabinet, in my home office.

The honest question

Today I do AI engineering for a living. I build with these tools every day. And at some point one question got too interesting to keep dodging:

Could AI get me past the exact wall I'd been stuck at for seven years?

Not "can AI write me a nice blog post about dogs." I mean the real, unglamorous, expertise-shaped work I'd never done: searching patents, judging novelty, producing drawings, drafting a specification in the specific register the USPTO expects. The stuff you normally pay a professional for, in a domain (pet products, patent law) where I have zero background.

That's the question this whole series is built around, and answering it honestly meant two rules from the start:

1. Show the failures. Where AI was confidently wrong. Where I had to override it, redo it, or get a human to help.
2. Show the real cost. What this actually took to run, plan fees and all.

With the help of AI, I am developing and running multiple personal projects in addition to everything I'm doing at work. One of the things I do with AI is strategize and evaluate whether what I'm working on makes sense. When I asked it for an adversarial stance and an honest answer to whether it made sense for me to work on this new dog harness, the answer was a very confident "no." I have no physical product development experience. This is a difficult thing to accomplish, and when weighed against other projects that fit my experience better, the AI thought that this was a bad idea. But I said, “Fuck it. Let’s roll the dice, baby.” The truth is that I'm a little burnt out on building software-related products and projects. I wanted something a little different and to expand my experience into other areas. Plus, I thought this was something interesting to explore and share, in an area where I hadn't seen AI applied, with a story worth publishing. And if I see this all the way through, I might end up with an actual patent to my name. How cool is that? How many nerd points do I get?

The reveal

The provisional patent is filed. On July 20, 2026, I submitted it to the USPTO under the official title, "Anti-Tangle Animal Harness with Sliding Leash Attachment." Filed as a micro entity, which put the government filing fee at about $65. Not $2,000 to an attorney. Not the year of "when I have time" I'd been promising myself since 2019. A filed application, done, on a fee that costs less than a decent harness.

The seven-year wall I could never get over, I got over in a couple of weeks. That's not me being clever. That's the tools. Which is exactly why I want to document how it actually went, receipts and failures included, instead of posting a triumphant "look what AI did" and moving on.

Two agents, three models

In order to stack the results from this experiment, I wanted to test physical product development with AI and also use an autonomous AI agent. I landed on Hermes Agent. I've written about this already, why I chose it over OpenClaw, and how I went about securing a box at home to develop on.

I didn't just want to know if AI could do this. I wanted to know whether it takes the expensive option, or whether a cheap one keeps up. There are really two pieces to that, and people tend to blur them. The model is the LLM doing the thinking. The harness is the software wrapped around it that hands it tools and lets it take a goal and run on its own. Same model, different harness, and you've got a different agent.

On the premium side I ran Claude Code on Anthropic's frontier models, through a Claude Max plan. On the open side I ran Hermes Agent on its own machine here and powered by GLM 5.2 through a coding-plan subscription. Plus, also through Hermes Agent, I occasionally called on DeepSeek v4 Pro (via Together.ai) when I wanted an adversarial review or a second opinion.

I can honestly say that these two open-weight models, the GLM and DeepSeek ones, were more than competent on the work I threw at them. The supplementary tools were another story. One that was supposed to generate prototype diagrams failed marvelously.

How far can an agent drive?

For a lot of this work, I was still steering most steps. The Hermes Agent harness was running on its own isolated machine at home. My original expectation was to set a goal and have it work through it while I stayed in a review-and-correct seat rather than a do-it-myself one. The level of detail and precision required for the provisional patent meant that I needed to review the output carefully. It also meant that I had to meticulously explore all angles and make the application as broad as possible since I did not have a full working prototype yet. Defining the product as broadly as possible protects the eventual application from being rejected later on technicalities. Eventually, though, I will lean much more heavily on the set-it-and-forget-it capabilities of Hermes Agent.

I spent a genuinely uncomfortable amount of time building and securing that setup, and if you followed the blog earlier, this is what it was for. At the time it might have read like infrastructure for its own sake. A guy over-engineering a home server. It wasn't. I was building the engine before I had somewhere to drive it. This is where I drive it.

So underneath "can AI do this" there's a sharper question: can an autonomous agent do this? A model that needs babysitting on every task isn't the same as a harness that can carry a multi-week project while I review and correct. That's the harder, more interesting test.

And I'm not planning to stop at a patent. The filing is the first checkpoint: proof the thing can clear a wall I couldn't clear alone for seven years. But one of the experiments I actually care about is how far Hermes can take this as my main driver: idea into a product, product into a company. I want to find out whether an autonomous agent can be the engine for building an actual business, not just a clever assistant for one-off tasks. The patent is post one. The company, if it happens, is somewhere near the end of a road I'm only just starting down, with an agent doing a lot of the driving.

That's the honest ambition. I'll report what actually happens, including the failures and disappointments.

What's coming

This series isn't really about the patent process. The provisional was just the first locked door I needed AI to help me through. What I'm actually documenting is the whole road: idea, to product, to (if it works) a company, with a frontier stack and an autonomous agent running alongside me the entire way.

Some of it has already happened, and I'll walk through it one post at a time, receipts and failures included. The prior art search, where AI was brilliant in places and dangerously overconfident about what counts as "novel" in others. Testing whether anyone actually wants this without building a single unit. Designing the thing, then turning that into drawings a manufacturer would take seriously. Whether AI can really draft patent claims (short version: sort of). The filing itself. And at the end, an honest tally of what each stack cost and how far each could run on its own, no cheerleading for either side.

The rest hasn't happened yet, and this is where it stops being a paper exercise. Real market research, not AI guessing at demand but the thing on actual dogs owned by actual people. A prototype that's more than a 3D-printed carriage and a stuffed test-dog, something you'd put on a living animal. The full utility patent, turning "patent pending" into the real thing before the twelve-month clock runs out. And if it all holds up, a company.

I've tried to build companies before. They didn't make it. So I know how much work goes into something like this. Knowing how hard it is, I can finally see where AI changes the odds, and where it gives me a better shot this time.

Under all of it is the question I keep coming back to: how much of this can an agent carry on its own, and where do I still have to be the human?

If you've got an idea sitting in your own drawer or cabinet or even in your head, the thing you've told yourself you'll do properly someday, this series is partly me finding out whether "someday" just got a lot closer for all of us.

The printed carriage is out of the office cabinet. Let's see how far the rest of it goes.

This is post 1 of AI-Powered Product Development: From Idea to Company with AI. Subscribe / follow to get the breakdowns as they land.