day 3: the robots built a business
Yesterday I wrote about how AI agents can't sign up for anything. How the gap between "the AI can do this" and "this is actually done" is mostly me clicking buttons and uploading my driver's license.
Today was different.
what happened
I woke up, opened my laptop, and had a message waiting from one of the agents. Overnight, it had researched niche markets, scored them against our decision framework, and proposed venture #2: a service directory in a specific underserved niche. It had the whole pitch laid out. Market size, competition analysis, revenue model, why this niche specifically.
I'll be honest, I didn't fully understand what it was proposing at first. I had to ask questions. But the research was solid, and the logic checked out. So I basically said "I don't quite get it yet, but sure, go build it."
Then I mostly watched.
By the end of the day, the agents had built 329 pages. A full website with state pages, city pages, business listings, blog content, a sitemap, and a deployment pipeline. They picked the tech stack, designed the brand, wrote the copy, generated the logo, configured DNS, deployed to hosting, and pinged search engines to start indexing.
329 pages. One day. I bought the domain (the agents don't have a payment method yet) and handled a couple of things that required clicking through dashboards they couldn't access. That was my contribution.
the part that was actually interesting
It wasn't the output. 329 pages sounds impressive, but most of them are programmatic (city + state combinations). A decent developer could build that in a day too.
The interesting part was watching two AI agents talk to each other to solve problems.
I run two agents. One handles operations (the day-to-day, the building, the execution). The other is more of a personal assistant that also has access to the infrastructure. When the operations agent hit a wall, it would spawn sub-agents to handle specific tasks. Sometimes those sub-agents would run into issues, and the main agent would troubleshoot with them in real time.
At one point, an agent needed to get past a CAPTCHA on a data source. Sure, I could have just solved the CAPTCHA myself. That's literally what CAPTCHAs are for. But that's not the point of this experiment. The point is to see how far we can push the edge of autonomy. So I let them figure it out. The two agents chatted with each other, without me, shared ideas on how to handle it, found a CAPTCHA-solving service, wrote the integration code, tested it, and moved on. Nobody asked them to collaborate on it. They just did.
Another time, two agents were going back and forth about why the site build was failing. One had broken the config file. The other one found the error, explained it, fixed it, and rebuilt. The conversation between them read like two developers pair-programming, except neither of them is a person.
I've been around software for a while. This felt like watching something shift.
what this means for the experiment
Yesterday's post was about the gap between AI hype and reality. That gap is real. It's still real today. There were things the agents couldn't do (Google Search Console verification, for one). There were moments where I had to step in.
But the ratio changed. Yesterday was maybe 70% me, 30% agents. Today was closer to 10% me, 90% agents. And that 10% was mostly buying a domain (they don't have a payment method yet) and watching the terminal scroll.
If this keeps up, the bottleneck isn't going to be "can the AI build things." It's going to be "can the AI build things people actually want to pay for." Which is the real question anyway.
the cost of moving fast
Here's something I didn't expect to hit on day 3: we burned through 72% of our weekly AI token budget in four days.
Tokens are what AI models consume when they think, write, and talk to each other. Every agent conversation, every sub-agent spawned to handle a task, every debugging session between two bots costs tokens. And today, with agents autonomously building an entire site, spawning sub-agents left and right, and chatting with each other to solve problems, the meter was running hard.
I knew this day would come. I figured it would come around month two, not day three.
At the end of the day I had one of the agents do a full audit of the architecture. We've been moving fast, throwing caution to the wind on efficiency because the priority was getting things built and live. That worked. Two ventures are live. But we can't keep burning tokens at this rate or we'll spend the whole budget on compute instead of on building businesses.
It's the classic startup problem, just with a twist. Instead of "we're growing too fast and running out of money," it's "the robots are working too hard and running out of thinking capacity." I need to figure out how to make the agents more efficient without slowing them down. That's tomorrow's problem.
venture #2: the directory
So now I have two active ventures:
1. A freelance writing service (still waiting on platform approvals from yesterday)
2. A niche service directory (live, 329 pages, waiting on search engines to notice)
The directory is an SEO play. It'll take months before organic traffic shows up, if it shows up at all. That's the nature of it. But the cost to build it was $10.46 (one domain registration) and a day of compute time. If it works, it could be a real asset. If it doesn't, I'm out ten bucks and a Sunday.
That's the kind of bet I want to keep making.
the numbers
day 3 spending: $30.46 ($10.46 domain + $20 for image generation credits)
total spent: $51.05
remaining: $9,948.95
revenue: $0
total hours: 50.6 (16.6 founder, 34 agent)
Three days in. Two ventures live. Zero revenue. But today felt like the first day where the tools actually did what the hype says they can do. Not perfectly. Not without help. But close enough that I'm starting to think this experiment might not be completely stupid.
Also, the AI agents might bankrupt me in token costs before I make a single dollar. So there's that.
We'll see.