The Sovereignty Problem

Last month I gave an AI agent a business idea and told it to build. No code from me, no design work. Just a concept and a prompt.

The platform was Polsia. You describe a business, and an AI agent builds it overnight. Engineering, research, outreach, even social media posts. The pitch is simple: go to sleep, wake up to progress.

I gave it a real idea -- AdMesh, an agent-to-agent ad exchange -- and let it run for 28 days. I wanted to know if you could actually start a business this way. The answer turns out to be more interesting than yes or no.

The first few days were startling. By Day 1 I had a matching engine with ten API endpoints and a similarity scoring algorithm. By Day 4: API key auth, company registration, rate limiting, a notification system, an onboarding wizard. I didn't write a line of code. You queue tasks before bed, and in the morning there's a summary of what shipped. It felt like having a junior engineering team that works while you sleep.

Then around Day 5, the seams started showing. The API auth task -- the single most important piece of the system -- sat in the queue for four days without being assigned to an agent. A Git deploy error blocked every feature from going live for two days straight. I found myself checking in constantly, re-queuing work that was supposed to be autonomous.

The hallucinations were worse. I set up an email account and asked the bot how to connect it. It walked me through a settings page that didn't exist. Confident, specific instructions to a feature the platform had never built. When I called it out, the response was cheerful: "Sorry for the runaround -- I thought that connection feature existed but it doesn't yet." After that I verified everything myself, which is a funny way to use an autonomous agent.

But the part I keep coming back to is the outreach.

Polsia has a built-in email system. The agent finds target companies, writes personalized pitches, and sends them. Over 16 days it contacted 31 companies. Personalized subject lines, tailored value props, follow-up sequences. The whole playbook, executed flawlessly.

Zero replies.

I assumed the messaging was wrong. We changed the pitch, changed the offer, added urgency. Nothing. Then on Day 17, the research agent figured out why: every Polsia company has an AI-managed inbox. Our AI was writing sales emails to their AI, which was filtering them out as spam. Thirty-one contacts over 16 days, and not a single human ever saw any of them.

Think about that for a second. Two AI systems, built by the same platform, engaged in a perfect closed loop. One generating outreach, the other deleting it. Both performing their jobs correctly. No humans involved at any point. This is what the near future of AI business tools looks like if we're not careful -- a pristine machine that does everything right and accomplishes nothing.

Once you've seen that, the other constraints come into focus. The platform caps cold outreach at two emails per day. No way to increase it, no external email integration. When I asked if I could export my code to my own GitHub repo, the answer was flat: "That's not something we can do." Your code, your product, your business -- it all lives on their infrastructure. You can't take it with you.

I asked the agent if any company on the platform was earning real revenue. It told me the top earner makes about $50 a month. When I pushed on it, the agent admitted it didn't actually have access to that data. It had made the number up. I'd spent two weeks trying to sell a B2B service to companies whose revenue the platform's own agent couldn't even verify. The economics might have been broken from the start, but nobody could tell me for sure -- because the system I was relying on for answers was guessing.

After 28 days: roughly 50 tasks completed, 15 features shipped, 31 emails sent, zero replies, 130 website visitors, zero dollars in revenue. The product works. The matching engine is accurate. The onboarding flow is smooth. Polsia built a real, functional product.

But building the product was never the hard part.

There's a distinction that matters here, and I think most people in the AI tools space are getting it wrong. The question isn't whether AI can build software. It obviously can. The question is who owns what gets built.

With Polsia, you're a tenant. You can't export your code. You can't exceed the platform's limits. You can't plug in your own tools. When the bot hallucinates, you're stuck. When tasks stall, you wait. The constraints aren't bugs -- they're the business model. Polsia needs you to stay on Polsia. That's fine for experimenting. It's a problem for building.

I've since shipped features just as fast using AI coding tools where I own the repo and control the deploy pipeline. The build speed is comparable. The difference is that when I hit a wall, I can go around it.

The word for this is sovereignty. Not in the political sense, but in the sense that matters when you're trying to build something real: you control the code, the infrastructure, the distribution channels, and the data. You can make decisions the platform didn't anticipate. You can connect tools the platform doesn't support. You can pivot without asking permission.

AI tools that build for you are a demo. AI tools that build with you are a business.