AI Journal 3: How to (not) vibe-code at work

2026-10-14

How to (not) vibe-code at work

My last AI journal entry was about vibe-coding at home for fun. I still do that. A lot. I will write more about that later.

At work, things have changed as well. Some big differences since I wrote my last journal entry months ago:

Let me address some of the points in more detail.

The cat is out of the bag

It seems everyone now know that agentic coding is a thing, and in many cases, people are starting to realize that the cat is out of the bag, pandora's box has been opened, etc. etc. It is no longer the secret of development team and people who code. I think there are many reasons for that. First of all, people outside of coding have started to realize how much stuff you can make AI do for you, like writing proposals, making flyers, mangle data, answering mails, and so on. When you tell them it can code as well, they believe you. The media is another reason. Stories like the "breach of Hugging Face" has changed peoples perspective on the capabilities of "AI". And also stories about how all the smart people think AI could destroy us all some day. The list goes on. I no longer have to explain to people that AI can actually code well. It's a given at this point.

Does everyone use "agentic coding" now?

I still have collegues who haven't actually tried agentic coding, but it's mostly by choice. In some cases, their job doesn't need it since they spend more time designing or reviewing than actually producing code, and in all cases I'm aware of, they still use chat-based AI for analysis and research. They just don't use it from their terminal (yet). I suspect they will have to adapt eventually.

Do customers require that we use AI?

Customers are starting to care about AI. I have yet to experience a customer who made hard requirements on how we should use it or which products to use. I suspect it's mostly because they just haven't gotten around to it yet. Also, as a policy, we are never going to promiss any cost savings because of AI. It doesn't work like that. Yet. Maybe we will be forced to lower our prices someday, but right now, no one is saving money because of AI, despite some things being done much faster than before. More on that below.

Telling "no" to the AI

One of the top stories on "hacker news" in the last 24 hours was Redesigning source control at Anthropic. I didn't actually read it all since I could not care less how Anthropic uses their own source control instead of git. But the comments on hacker news are always interesting. One comment highlighted the following quote from the story:

Very often, we found that the role of the engineers was to tell Claude “no.”

Yes, exactly! Even with the most recent models (I'm currently bouncing between "Opus 5.5" and "gpt-sol-6.1"), I feel like I have to spend a considerable amount of energy trying to minimze the scope of the code it produces. I've become quite good a prompting the AIs to not produce too much code, but sometimes, it still slips through. Just today, I caught Claude adding transactions and locks to a batch job in one of our services -- all I had asked it to do was switch the data model from "V1" to "V2" (which is essentially just adjusting the queries to a newer set of tables in MariaDB). And there was nothing wrong about the transaction manager or the locks -- in fact, they are used already somewhere else in the service, but it was out of scope for this specific change for various reasons. The point is, the default for most models is to do "too much", and that goes for all the models all the time. Restraint is still not part of the training, it seems.

Cleaning up after the AI

Another down-side of agentic coding is collegues that use it too much. Everyone has heard the stories of the PRs with extensive changes to the whole code base just because. Reviewing that is not fun. Luckily most of my collegues are sane people that don't do that. But there is another down-side to having inexperienced collegues who use AI, and that is when they start to tackle tasks that are way over their head. In the old days, they would give up, ask for help, decline to start the task in the first place, etc. Today, they will force their way through, and delegate the hard technical and architetural decisisions to claude or codex. The PR will look fine on the surface, not too big, all tests green, and so on. But when you start to dig in, you will find hidden assumptions that don't correnspond to business needs, and small refactors that change the existing business logic in subtle ways. And then you have to explain in the PR what your concerns are. And then they will paste those concerns into the AI product of choice, rinse and repeat. At this point, they are becoming a "meat proxy"

Has AI made us faster, cheaper, better yet?

I really can't tell if AI is a net benefit at this point. But I can speak for myself:

This is all great and I am happy.

I've already mentioned some of the downsides, but just for completeness sake:

So that's all for know. More journal entries to follow.

← Back home