Post

So, I've been vibe coding this event website that's taken a couple of turns here and there as the details about the event change. I've deliberately avoided looking at the actual code any more than necessary.

I've used several different agents so far based on my intuition of which excel at different tasks (e.g. Gemini 3 for creative work, Opus 4.6 for business logic, GPT-5.2 for copy and i18n).

So far, I've been way, way more productive than if I'd tried to build the thing on my own, or even with one or two other people. However...

Now that the project's reached a certain level of maturity and been worked on by multiple agents, some real fault lines are starting to appear.

I've noticed that the agents are taking longer and longer to get oriented in the project and that makes sense. The code's unreadable. Every time they start on a new task, they have to learn how the project's organised from scratch.

The don't care at all about code reuse. I've found the same utility function written different ways in literally five different places.

And this is just a simple event website. If it was a serious business application, I'd now have to spend a lot of time figuring these issues out and either fixing them myself or letting an agent do it, but looking over its shoulder very, very carefully.

Just saying, this is amazing and everything, but I think we're still a long, long way from being able to one-shot prompting real software.