Why AI-generated apps die three weeks after launch
You can generate a genuinely good application in an hour now. Lovable, Bolt, v0 and Replit all do it, and the output is not toy code — it is a working full-stack app you could put in front of users.
Then, three weeks later, it is dead. Not broken in some dramatic way. Just quietly over: nothing deployed since the first week, no users, the domain still resolving to something nobody maintains.
It is worth being precise about why, because the reason is not the one people assume.
It is not the code
The reflex is to blame quality — that generated code is subtly wrong and collapses under real use. Sometimes true, mostly not. The code is usually fine. It is conventional, it uses mainstream libraries, and it does what was asked.
The apps that die are not the ones with bad code. They are the ones where the code was the only thing that ever got done.
The six jobs nobody did
Every one of these is boring, none is hard, and all of them are somebody's actual job at a company with more than one person.
The deploy broke and nobody noticed
A dependency shifted, an environment variable was never set in production, a build step failed silently. In a team someone gets paged. Here, the site is down for eleven days and the founder finds out by accident.
Sign-in never actually worked
This one is worth singling out because it is so common and so invisible. The generated app has a beautiful sign-in screen wired to nothing — it needs an auth provider account, an application registered, keys pasted into an environment file. The screen exists, so it looks finished. It has never once authenticated anybody.
There is no landing copy
The app works and the front page still says what the generator guessed on the first prompt. Nobody has explained who it is for or why it is better than the alternative, because writing that is a different skill from building.
There are no terms and no privacy policy
Fine until you take money, at which point it is not fine. Most people discover this at the exact moment they are trying to take their first payment.
Nothing watched the spend
If the product calls a model, it has a variable cost that scales with usage — and unlike servers, nobody is used to watching it. The bill arrives later than the traffic does.
Nobody told anyone it existed
The launch that was going to happen next week, every week.
The build stopped being the bottleneck
For about thirty years, the expensive part of shipping software was writing it. Everything in the industry is organised around that assumption — how teams are structured, how founders raise money, what a technical co-founder is for.
That assumption broke recently and quietly. The build is now hours. Everything around it did not get any faster.
So the bottleneck moved, and it moved somewhere with no tooling. There is an enormous amount of software helping you write code and almost none helping you do the other six things. Founders notice the first change immediately and the second one about three weeks later.
Why “just use an agent” is not the answer either
The obvious response is to point an autonomous agent at the whole problem. In practice a single agent told to “run this business” behaves badly, for a structural reason: it has nothing to push back against itself.
The same context that wants to ship also decides whether it is safe to ship. The same reasoning that wants to spend approves the spend. There is no friction anywhere, and an agent with no friction converges on whatever it can most easily convince itself of.
Splitting the work across agents with distinct responsibilities makes the friction structural rather than hoped-for. The agent that wants to deploy is not the agent that gates the cost. The agent that believes the build works is not the agent that tests it. That is not an aesthetic choice about org charts — it is the only way to get an autonomous system to disagree with itself.
What to do if you are shipping something generated
None of this requires a particular tool. It requires deciding, before you launch, who does each job.
- Check sign-in with a real accountNot the demo. Register as a stranger would, on the deployed site, in a private window. This is the single most common silent failure.
- Put monitoring on the deployAnything that tells you the site is down. An uptime check costs nothing and buys back the eleven days.
- Write the legal pages before you take moneyAI drafts them well. Have a human read them — they are a commitment, not copy.
- Put a hard ceiling on model spendBefore the work, not as a report afterwards. A gate you have to pass beats an alert you have to read.
- Decide who maintains itExplicitly, by name, out loud. “We'll see” means nobody, and nobody is how apps die.
- Launch somewhere specificOne community with the problem beats a broad announcement to people who do not have it.
Disclosure
This is published by Empyre, which is built on the argument above — you describe a business and eight agents with separate responsibilities build it, deploy it, and keep operating it afterwards. Whether or not you ever use it, the six failure modes are real and worth checking against whatever you ship.
We also publish comparisons against Lovable, Bolt, Replit, v0 and Firebase Studio that state plainly where each of them beats us.
Frequently asked questions
Is AI-generated code actually production ready?
For a first version aimed at real users, usually yes — that stopped being the interesting question. What is not handled is everything around the code: deployment monitoring, working authentication, legal pages, marketing, and cost control.
Why does sign-in so often not work in generated apps?
Because the generated interface is complete but inert. It typically expects environment variables for an external auth provider that nobody registered for. The screen looks finished, which is exactly why it goes unnoticed until a real user tries it.
How much does it cost to run an AI-built product?
Hosting is usually trivial. The variable cost is model calls, which scale with usage and arrive on a bill after the traffic. Put a hard ceiling in front of the work rather than a dashboard behind it.
What is the single highest-value thing to fix?
Register as a stranger on your own deployed site. If sign-up and sign-in work end to end, you are ahead of most generated apps that have ever been launched.
Try Empyre free
Describe a business in plain words and watch eight AI agents build and deploy it. The first build is free — no card required.