June, a Sunday, and the night before it.
I had promised the class their own coding agent.
The reasoning was simple enough. The good systems are expensive. I was not going to ask people who are retraining, most of them paying for this out of savings, to also carry a bill that grows every time they practise. So I decided to stand up something cheaper and give them access to it.
I worked on it until one in the morning.
The plan was reasonable. Take one of the open systems, the kind you can run on infrastructure you choose rather than renting through somebody's front door. Host it somewhere inside the United States, because the first question any employer asks about a cheaper option is where the data goes, and I wanted an answer that survived that question. Wire it up so the students could reach it from their own editors.
The pieces existed. There was a provider selling access on domestic infrastructure. There was a standard way to speak to it. I had done harder things.
And it nearly worked, which is the worst outcome, because nearly working keeps you at the desk.
The problem turned out to be sitting between the pieces. The provider was translating my requests into the shape the system underneath expected. For ordinary conversation that translation was fine. But when I asked it to drive a browser and run a real test, the request went through that layer and came out slightly wrong. Not broken in a way that produced an error I could read. Wrong in a way where the thing simply did not do what I had asked.
I sat there past midnight taking it apart, and at some point I understood that I was not going to have it ready.
The next morning I told the class exactly that. I said I worked until one and I could not finish, and here is what stopped me, and this is what the work actually looks like. I showed them the terminal. I showed them the thing half working.
I want to be clear about why I told them rather than quietly pushing it a week. It was not humility as a performance. It was that they were about to walk into a market where they will be asked to do this, and the version they see online is a demonstration where everything connects on the first try. I have never once had that experience. The value of the morning was not the working agent. It was watching somebody competent fail at an integration in front of them and describe the failure in plain language.
The part of that class I still think was worth more than the agent came earlier, when I explained why I was doing it at all.
I asked them what happens to a software company when it adds a customer. For the last twenty years, the answer was almost nothing. You build it once, and the ten thousandth customer costs you close to zero. That is why software became the best business anyone had ever seen.
Then I asked what happens when the product thinks. Every question costs. Every answer costs. The ten thousandth customer is not free, and neither is the same customer asking twice.
You could watch it land. Somebody said, so the margin is gone. Not gone, I said, but it is not what it was, and every company selling you these tools is currently paying part of your bill to get you in the door.
I told them how that story usually ends, because it has happened before with the cloud. Come in below cost. Get everybody's infrastructure sitting on top of you. Then reprice, once leaving costs more than staying.
Somebody asked what you do about it.
I said you become the person who knows which work needs the expensive thing and which does not. Because that person saves the company real money, and saving real money is the easiest salary conversation there is.