Look What I Built

I want to see whether it has sorted out the login. Before I let the agent rewrite anything, I check whether the app is even pointing to the right Firebase project. There is no point changing the lock if you are standing outside the wrong building. It should take five minutes.

The next time I look at the clock, it is two in the morning. I cannot really account for what I have spent the past hour and 45 minutes doing. I also cannot account for what the agent has done until I have read its changes. Right now I am staring at a cursor blinking next to the word Thinking.

Then comes the ding. The build is green. The app can be built without errors, and the agent tells me the solution is better than before. I have not even tried to log in yet, and I am already in the mood where I would quite like someone looking over my shoulder. The thing is, nobody is waiting for this product. If there were a waiting list, I would be both at the top and at the bottom.

I think I have got myself a Tamagotchi again. If you are not old enough to have had one: it was a tiny electronic egg from 1996 with a digital pet on a screen. You had to feed it and clean up its mess, and if you ignored it long enough, it could die. My code agent can manage without me. It does not even need to beep. I go in and check on it myself.

It is almost worse when it works. If the login works, we might as well get the invitations working too. And when those work, we will presumably need a page where you can see who has accepted them. There is still nobody to invite. But now I can see that they have not replied.

In Meet the Sad Wives of AI, Alessandra Ram tells the story of her husband, who calls from a hotel room at two in the morning to show her something in Claude Code. She is at home with their daughter. The dog still has to be taken out. Several of the women in her report describe men who are afraid of missing the shift they believe will define the rest of their working lives. Meanwhile, someone still has to hold the everyday together. I know which side I am on in that scene. I also know why he calls.

I am myself thoroughly fed up with hearing men explain what they can do with AI. Especially the ones who cannot talk about anything else. It is an opinion I find easiest to voice when someone else has the computer open. In the World Economic Forum's 2026 report, women made up an average of 19.3 percent of AI engineers across the 19 economies studied, according to LinkedIn data. Roughly four out of five were men. That is a specific professional group, not all of us who say we work with AI. It is, incidentally, a wonderfully broad phrase. It can mean both that you develop a language model and that you have asked Claude to build a rocket out of papier-mâché. I would like to see both. I just do not want to be sitting in the latter.

I work with AI every day, and I would not want to be without code agents like Codex and Claude Code. They can change files, run tests, and help me try out ideas I would otherwise have parked. I can also put the code helmet back on afterwards. When I review a login, I am just as interested in what a user does not have access to as in whether they can get in. You cannot see the first in a screen recording of the second.

Even Anthropic, the company behind Claude, has looked into the difference between getting code to work and understanding it. In an experiment with 52 developers, most of them relatively new to the field, the group with AI assistance scored an average of 50 percent on the subsequent test. The group without scored 67. The biggest gap was on debugging questions. It was a short task with an unfamiliar Python library and an AI assistant, not a test of Claude Code. But the researchers were looking at something that rarely makes it into a demo: How much did you yourself understand of what you just got working?

I have looked under the bonnet of enough shiny prototypes to know I would not entrust all of them with my first name. What data does the solution have access to? What does it pass on? What happens when the service it depends on does not respond? Is it still cheaper once you factor in oversight, fixes, and operations? You do not have to have written every line yourself to answer that. But you also cannot just ask the same agent that has just told you everything looks splendid.

And what exactly have we become faster at? If Jytte in marketing can now send 42 pages instead of four, and nobody read the four, the only certain progress is that there are now 38 more pages nobody reads. That mistake is easy to spot in Jytte's report. It is harder to spot when you chose the font yourself.

Recently I saw an Instagram story making a point roughly like this: Have you also spent a thousand hours on an interface no one will ever use? Then you will be sorry to hear that the next big thing is no UI: you ask the software for what you want done instead of hunting for the right button. I do not know whether we are done with buttons. I only know that I have spent an indecent amount of my adult life tending to some that no human being will ever press.

Show Me the Rest

I think part of what makes it so hard to stop is that, for one evening, I get to feel like a man who can build anything. That feeling is real enough. It is the expertise I may end up claiming in advance. If I just want to see whether something can be done, I have carried out an experiment. That is enough. The moment I ask other people to use it, I owe them an explanation of what I am asking them to trust.

The next time someone says, "look what I built," I will still want to see it. Show me what happens when something breaks. Tell me who would miss it if you switched it off. I can handle the first demo. I can explain how my login works. The hardest question is still who the hell would want to log in.