I’m testing how far AI can go building real apps — not just landing pages

I've been building DreamAgent, an AI development platform that turns natural-language requirements into working applications.

Instead of testing it only on landing pages, I've started pushing it with real projects.

So far I've built:

☕ A premium coffee ecommerce website

- Product catalog

- Product cards

- Cart flow

- Responsive UI

🤖 A crypto Discord bot connected to Binance

- Live crypto prices

- Market data

- Price history

- Charts

- Discord commands

The interesting part isn't getting the first version generated.

It's what happens afterwards.

I can continue giving the AI instructions to modify the application, add functionality, integrate APIs, and fix issues.

For the Binance bot, I also tested using API credentials through environment variables rather than hardcoding secrets.

I'm now testing the same workflow with a Telegram crypto bot.

My current question is:

How far can AI-assisted development realistically go before a developer needs to take over?

The areas I'm watching closely are:

- API integrations

- Authentication

- Environment variables/secrets

- Debugging

- Maintaining existing code after multiple AI edits

- Deployment

- Complex feature changes

I'm sharing the experiments publicly because I want feedback from other indie hackers who are actually building products with AI.

If you've used AI coding/app-building tools for a real SaaS or product, what was the first point where the AI stopped being useful?

I'd especially like to hear about problems you encountered after the initial prototype stage.