It works. I just don't know why.
Contents
Claude Code has become enormously capable in a short time. It stopped being only about writing code a while ago. With browser and computer use, it also finds its way around another vendor's interface and sets up whatever my integration needs there.
When I want to add something new to my system and need another provider's API for it, one sentence is very close at hand: "Take care of this for me."
Most of the time, that works. And that's what I've been thinking about for a while now.
A success that isn't quite mine #
I'm used to programming myself and working solutions out on my own. But next to a job, a family and other commitments, there isn't much time left for hobby projects. An agent that saves me an evening of configuration is a real relief. I get to results quickly, and I don't want to play that down.
Then everything works, and I'm sitting in front of it. I didn't do the configuration myself. I don't know why it works. In a way, I feel robbed of something: the sense of having earned it and learned something along the way.
When I code with an agent, I at least still have the diff. With browser and computer use, even that goes away. What gets set in a vendor's interface doesn't show up in any commit. The result is there, but I can't read back how it got there.
My judgment comes from the effort #
If it were only about me, I could live with that. When an agent suggests something odd, I usually notice. But I only notice because I had to work these things out myself for a long time. My judgment comes from exactly the effort I'm now handing off.
So I'm living off a reserve. It no longer grows by itself, and without practice it shrinks. I can tell, because I know what it was like before.
Someone starting today doesn't.
What happens to the next generation? #
Anyone learning software engineering right now has a tool from day one that solves almost every exercise. Why spend two hours on a bug when the fix is two minutes away? I wouldn't have resisted either. This isn't an accusation against a generation. It's reasonable behavior, just under the wrong incentives.
The calculator comparison is an obvious one. Children still learn mental arithmetic, so they notice when a result can't be right. That part holds. But a calculator gets the math right. An agent also produces wrong solutions, and they sound just as convincing as the right ones.
People who only ever work with finished solutions eventually lose the ability to check them. Then nobody checks them. And these are the people who are supposed to become the senior engineers overseeing agents later on.
Understanding becomes a decision #
I don't think abstinence is the answer. Training without agents prepares people for a working world that no longer exists in that form. But something else has shifted. Understanding used to be a by-product of the work, because there was no other way to get to the result. Now it's a decision.
For training, I see three consequences.
The fundamentals come first and by hand, even if that's slower. At this stage, the detour is the content.
What gets assessed is whether someone can explain why something works. The fact that it works no longer counts as proof. "Why this way?" becomes the most important question in a review.
Agents still belong in training, as something to be critical of. Learning to doubt their output and find their mistakes trains exactly the judgment this is all about.
For myself, I have a simpler rule. Before I start, I ask whether I just want to have this or also want to be able to do it. If I just want to have it, the agent can take care of it. I don't need to master every setting at every vendor. If I want to be able to do it, I do it myself. Or I have the agent explain afterwards what happened, and I change something by hand. Anything that keeps running on my side and carries responsibility, I have to be able to explain either way.
An agent can take the work off my hands. It can't do the understanding for me. If you don't force yourself to learn it, the competence goes, and you only find out on the day the agent is wrong.
I wrote about how I set limits for agents on long-running work in My agent had no stop condition.