As code gets cheaper, verification gets more valuable

Something I found interesting in OpenAI’s latest report on how its researchers are using coding agents.

AI is making the implementation part of engineering much easier. Researchers are writing more code, running more experiments and taking on increasingly complex tasks with agents.

But there’s an interesting shift happening underneath that.

As the cost of producing code falls, verification becomes more important.

OpenAI found that even among successful 4–8 hour tasks, more than half involved at least one human intervention. And the more complex the task, the more human steering agents still needed.

The valuable skill may not be writing the code yourself.

It may be knowing what good looks like, how to test it, what edge cases to look for, and when the output should not be trusted.

AI can make creating software much easier.

Making sure the software is right is a different problem.

And as AI gets better at building, I think engineering judgment becomes even more valuable.