You don’t need Quality everywhere.
You only need Quality where it matters.
It doesn’t matter if your website is of Quality; in the past, we would outsource the development to a freelancer and never peek at the code. As long as it loads fast, look reasonably good and functions, there is no complaint to have.
It’s the same thing with GenAI; honestly who writes their frontend themselves anymore? If that makes you happy, then sure go ahead; if you really need quality (i.e. having it load 1s faster would really increase revenue), then yes, go ahead and work on it yourself.
If you predictably know that it doesn’t matter, then in my opinion, it’s fine to delegate the work to a freelancer, or in 2026, to GenAI.
In my opinion, we should be focusing on putting boundaries
There could be two categories of repositories:
- This piece of software is critical, thus, AI code shouldn’t. be let in without thorough review. GenAI can be used as a helper, or for drafting code, but it should never edit the files directly. The human is responsible for typing it, and making judgements regarding Quality.
We don’t mind if it takes more time, because Quality here is useful.
This is the “romantic” view.
- This piece of software must work, but Quality doesn’t matter.
It’s a piece of work that:
- Is not critical
- Could be easily delegated / swapped
- Can be achieved by GenAI with minimal steering
This is the “functional” view.
The resurgence of microservices
So, what does that lead us to?
Microservices.
What if you just asked the AI: get me that data, you figure it out.
You have your own repo, you’re responsible for monitoring it.
Sure it’ll be ugly, somewhat unreliable, and might produce incorrect data.
However, it can be a really nice placeholder until we get time to rewrite it with Quality in mind.
I think a lot of people started thinking that monoliths are the way to go because it allows agents to easily access context; you can define an AGENTS.md at the root that allows defining rules for the whole company.
What if instead, we had one central repository with Protobuf definitions, a little diagram, and just let each agent build ZeroMQ endpoints?
I don’t really mind how my agent fills my Protobuf payload as long as it’s following the definition.
What matters to me though, is having a broker in the middle that will pull the work that these AI-generated repositories create and then validate that the data makes sense.
It matters to me to be able to evaluate the quality of the output (but not so much the quality of the code that creates it).
But I’m 100% fine with a blackbox that does it’s job in the same way that I would be fine with a coworker with different tastes; as long as I’m getting the data I need from their API / lib and they don’t get in my way, that’s fine.
Everyone can have their little island of sanity; let the LLM struggle with their own insanity in their own repository!
I’m not sure if it makes sense; I’m not the best writer either; I just wanted to get this out of my head and maybe send it to a few friends.
If this resonates to you, maybe read this post.