AI can write code that works. That’s not controversial anymore.
What it struggles with is code that survives.
Most AI generated code fails in production for a very boring reason: it assumes the happy path. No retries. No idempotency. No partial failures. No weird states.
In real systems, networks flake. Jobs run twice. Requests time out halfway through. That’s where things break.
The AI isn’t “bad” here. It’s optimizing for helpfulness, not durability.
This is why concepts like the twelve factor app still matter.
https://12factor.net
Or why people like Martin Fowler still talk about boring things like boundaries and responsibility.
https://martinfowler.com
The fix isn’t banning AI code. It’s reviewing it with a production mindset. Asking boring questions like: what happens if this runs twice, what happens if this fails halfway, what happens under load.
If you can’t explain those answers, you’re shipping hope, not software.
Top comments (0)