The lab that shipped itself

The paid edition of my own AI lab launched this weekend. The strange part is how it got built. The lab shipped itself.

That sounds like a riddle, so let me explain. Arena is the free cockpit I use to run every model and test I make on one desktop, the NVIDIA DGX Spark. The Field Edition is the paid version: one installer that stands up the whole stack on a Spark you own, then proves it works. To launch it I needed a store, prices, a way to deliver a license, and the install itself. And the tool I used to do that work was the same Arena the launch was for.

The plan said ten weeks

My strategy agent drew up the launch plan first. It was a good plan, careful and staged, with the public date set about ten weeks out, late in August. That is the honest number a sensible plan returns for a first paid product. A store, a license system, fulfillment, and a clean install are all real work.

I read the plan and made a different call. The code was closer to ready than the plan assumed. So we shipped the whole thing, store and license delivery and all, in a single weekend day. The part the plan had penciled in for ten weeks took one day!

The team I conduct

Here is who “we” is, because not one of them is on a payroll.

On my Mac, three AI agents. One works on strategy and keeps the plan honest. One works on marketing and writes the launch. One designs and builds this website, the page you are reading right now. On the Spark, a fourth agent works as a research engineer that never sleeps. It runs the lab: training models, scoring them on tests, writing the install.

I do not write the code anymore. I conduct. I set the goal, settle the hard calls, and pass work between the agents like a conductor cueing sections of an orchestra. The strategy agent hands a brief to marketing. Marketing asks the website agent for a page. The website agent asks me to decide a price. I point the research engineer at the next build. The work moves in a loop, and I keep the beat.

The lab shipped itself

This is the part I care about most. The word for it is dogfooding, which just means using your own tool to make your own tool. Every step of the Field Edition launch ran through Arena’s own screens. The models it sells were trained and scored in the eval drawer. The serving was swapped on the guarded model pane. And the install ends the way the product promises: it proves itself on first boot and writes a verifiable receipt, a small report that says what was checked and what it found on your box.

So the cockpit did not just record its own launch. It was the instrument the launch happened on. A tool that can ship itself is a tool I can stand behind.

A day, measured

I do not want to round this up, so here are real numbers from the lab’s own logs. About 570 code changes from March to June, with sprint days running fifteen to forty of them. One model, Advisor, went from idea to a public, tested product in about thirty hours. The toolbox under all of it once shipped sixteen versions in six days. A study of the year called the output what a five to eight person team would produce. The plainest line in it: the cost of a new product here is days, not quarters.

The reason this is affordable is dull and important. The agents reread their own past work from a cheap cache about 97 times in 100, so each new step pays only for the new thinking. That is how one person keeps a team of agents busy without spending a fortune.

What you actually get

The pitch is simple. Get an AI team without hiring one, inside the DGX Spark you already own.

The proof is the shelf this exact setup already built, solo on one Spark: 14 software tools, 6 open models, 3 books, a 54-article research library, and two production websites, this one and ainative.business. That is the velocity the box unlocks.

The software stays free and open. You can run the free Arena today, and there is a live demo you can click through without installing anything. What you buy with the Arena Field Edition is the assembled team, delivered turnkey: agents that do real work, models that know your field, and tests that prove it. Not seven free parts you wire together and babysit. One installer sets it up, it proves itself on first boot with that verifiable receipt, and it stays proven for a year, every update re-checked and re-receipted.

It is $349 for the first 25 founding licenses, then $499 one time, with a $149 a year renewal after the first year if you want to keep getting fresh proof. The same team that shipped this in a day is the team you are buying. That is the whole offer.