Loop Generation? A Programmer's Dream, Maybe

The Idea of Loop Generation

Recently, I have started seeing the term “loop generation” more often.

It is not just about giving AI a prompt and waiting for the output.

The idea is that AI looks at the result, decides the next task, performs additional research if necessary, and repeats the process.

In short, it is about making AI operate more autonomously.

The use cases people mention are easy to understand.

Meeting note cleanup.

Email checks.

Regular market research.

News collection.

These are routine tasks that happen every day.

AI is already quite good at that kind of work, and those tasks fit well with loops.

Humans no longer need to give the same instruction every time.

There is no doubt that this will be convenient.

What Can Be Turned Into a Loop

But as someone who has been a programmer for a long time, I also see it a little differently.

Work that can be turned into a loop is, by definition, work that has already been simplified.

The input is defined.

The procedure is defined.

The goal is defined.

That is why it can be repeated.

That is why it can be automated.

That is why it becomes a loop.

When AI Research Goes Wrong

The other day, I asked AI to handle a research task.

It produced a very impressive document.

The structure was clean.

The writing was natural.

The problem was that the content was almost entirely false.

And the AI itself showed no hesitation at all.

It was completely confident.

I almost believed it too.

After experiences like that, it is still hard for me to imagine development work being run as a loop.

Development Is Mostly Different Every Time

Software development is different every time.

The problem to solve is different.

The constraints are different.

The history behind the system is different.

The amount of information and the level of complexity are both high.

Almost every time, we are facing a new problem.

Of course, there are simple parts within development work too.

Code reviews.

Design reviews.

Log analysis.

Documentation.

Research.

AI has become quite good at those things.

In fact, I use Claude Code Skills a lot myself.

If I turn review perspectives or research procedures into Skills, I do not need to explain the same things every time.

As a way to make routine work more efficient, it is extremely powerful.

The Core of Development Is Elsewhere

So I do not think loop generation is useless.

On the contrary, I think work like meeting note cleanup and regular research will increasingly be handled through loops.

But that is not the core of development.

The center of development is the part that is hard to turn into a Skill and hard to place on a loop.

In the end, what remains is exception handling.

That is why I am not rejecting loop generation itself.

I actually think we should hand more routine work over to AI.

But if software development itself can truly be run as a loop, that is a different story.

At that point, we will no longer be debating how to use AI.

We engineers will be asking AI how to get to the employment office.