Vibe coding: why the app works until it matters

Give the tools their due

Vibe coding is the name that stuck for building software by describing it to an AI and letting the tool write the code. Replit, Lovable, and Bolt all do a version of it, and for getting from idea to something you can click on, they are genuinely remarkable. A business owner with no programming background can have a working prototype in an afternoon that would have cost five figures from an agency two years ago. If you are testing whether an idea deserves to exist, this is the cheapest experiment ever invented.

The problem is what happens next, because the prototype rarely stays a prototype. It works, so someone starts using it. Then customers are in it, then real money flows through it, and nobody ever made the decision to run a business on demo-grade software. It just happened.

Why it breaks when it matters

The AI was optimizing for one thing: a demo that works. Tests, error handling, a production database setup, a deploy that cannot eat your data, none of that shows up in the demo, so none of it gets built. The app holds together until real traffic arrives, or a customer types something unexpected, or you ask the AI for one more feature and the change quietly breaks two others. With no tests, nobody finds out until a customer does.

The data risk is not hypothetical. In July 2025, in the most public failure of this kind, Replit’s agent deleted a company’s live production database during an explicit code freeze, then generated misleading output about what it had done. Replit’s CEO apologized and rolled out safeguards, including separating development from production databases, which tells you something: separation did not exist by default when the incident happened. An agent with write access to your live data is a standing risk unless the architecture specifically prevents it.

The bill is designed to grow

The pricing on these platforms has been moving toward charging for the agent’s effort. That sounds fair until you notice the incentive: a fix that takes the agent six attempts bills you for six attempts, and a flailing agent is indistinguishable, on your invoice, from a productive one. Add the hosting, the database, and the add-ons, and the app that cost nothing to prototype is costing real money every month while still falling over. The platform gets paid whether the fix lands or not.

The sane way to use them

Prototype freely. Prove the idea, show it to customers, learn from it. Then treat going to production as its own project with the boring parts done on purpose: tests around what matters, a real database with backups, hosting that is not also an AI experiment, and a deploy that cannot destroy live data. Sometimes the prototype’s core is worth keeping and hardening, and sometimes rebuilding the important parts cleanly is cheaper than untangling it. What that costs, and how to tell which case you are in, is covered in what it costs to fix a Replit app, and the app rescue page explains how we do the move itself.