On 14 September 2025 I published a short essay saying that AI bends the iron triangle but does not break it. Speed stretches, budget stretches with an asterisk, quality stays human. That month I made zero commits to any repository. I had a day job I was doing badly, a household I was doing badly, and a company I had started on the side and could not find hours for. The essay was written by someone with no time, and it reads that way now: “quality stays human” was less an observation than the author holding on to the one thing that was still his.
A year later I have numbers to check it against. That is the other change. In 2025 I wrote down what I felt about the work. In 2026 I can pull what the work recorded about itself: the commits, the tokens, the hours, the hour of the day. A reflection with a number under it is a different kind of sentence. It can be wrong, and next year I can find out.
Here are the numbers, and here is what they say.
What changed, in three moves
I did not get faster. I changed three things about how work reaches me.
Nothing gets built before it is written down. Every change, in every repository, starts as a proposal: what problem, what the outcome looks like, what is out of scope. The proposal gets a ticket number, the ticket number becomes the branch name, and the build happens after the thinking rather than instead of it. I use OpenSpec for this; the tool matters less than the rule.
I built a workshop instead of buying tools. Fifteen small servers that speak the Model Context Protocol, one per system I depend on: my notes, my calendar, my invoices, my tasks, my writing, the weather at home. On top of them, four companions: two for the engagements I carry, one for training, one for the day itself. A companion is not an agent. It reads, it drafts, it proposes; I decide and I send.
I stopped selling hours. Work is quoted as a fixed price for an outcome, from a spec the client has read. If the method makes me fast, the speed is mine. On an hourly invoice it would have been a discount.
The numbers
Everything below comes from my own systems, pulled on 9 September 2026.
- Commits. September 2025: zero. Since May 2026: between 900 and 1,400 a month; 1,338 in August.
- Tokens. August 2026, one machine: 6.9 billion tokens through my coding sessions, 22 million of them written back to me. At list prices, about 6,800 dollars for the month. I pay a flat subscription.
- When the machine runs. 65 percent of that work happens between 09:00 and 18:00, 23 percent between 18:00 and 21:00, 9 percent after 21:00. 87 percent of it on weekdays.
- Hours. Since I started tracking in mid-June: 614 hours across the day job, the company and the house. Twenty-nine of those days ran past ten hours.
That last line is the one I did not want to write.
Speed walked across the table
I wrote that speed was no longer the bottleneck. What I did not write is where the bottleneck went.
This summer I built a client’s product and had it finished on 21 August, waiting for content. On 9 September it was still empty. Nobody was slow. The people who own the content have day jobs and families, and the project now waits on their calendars instead of mine. A year ago I was the bottleneck. Now the bottleneck is other people’s attention, and no amount of tokens moves it. That is not a problem to fix. It is the schedule finally belonging to the right people.
Quality moved upstream
Quality did stay human. It changed address.
In 2025 the human part was downstream: reading the output, checking it, catching the subtle errors. That was the hidden work I named at the time. In 2026 the human part is the proposal. If the problem is written down well, the building is the boring part and goes right. If it is written down badly, the build is fast and wrong, and no review catches what the spec never said.
The hours did not shrink. They moved to the beginning of the week, and to a table with the person who carries the problem.
Budget, and the line I did not want to write
The subscription is flat, so the budget corner stretched. But the cost I was worried about in 2025, the checking, is not where the cost went.
Twenty-nine days over ten hours in twelve weeks is where it went. The tools gave me hours back and I spent a good share of them on the tools. Some of that is craft. Some of it was an evening that was supposed to be someone else’s. A freed hour does not stay free. You decide what it is for, or it fills with more of the same.
What actually moved
In September 2025 I could not do a good job at work, at home, or on the side. In September 2026 I am never late and never stuck in a list. At the day job I have hours to think, to prepare, to let a mail be read twice before it goes out, to be in a meeting rather than half in it. The community sites I look after get done with ease, and I have room to think about what I want to do for them rather than how little room there is.
Everything moved to the other side of the table. The waiting, the deciding, the content.
One more thing I did not expect. The most useful brief I received this year came from a colleague I would not have chosen to hear it from. They were the first to point at something in how I work that annoyed them, and they were right, twice, on two different mails. In 2025 I ended the essay by saying that briefing well is a form of respect. This year the respect ran the other way, and I took it.
A year later
Does the triangle still stand? For me, no. The triangle was a choice between three things I could trade against each other. There is nothing left to trade. Speed comes from the tools, and I no longer pay for it in hours. Quality is decided in the document I write before anything is built. Budget is a flat subscription on my side and a fixed price on the client’s.
What is still negotiated is scope, and it is negotiated once, in that document, before there is a price. After that the price follows from the scope, the quality follows from the document, and the speed is mine. The one thing I no longer control is the calendar. That belongs to the people who have to decide, and they decide in the hours they have.
Nearly every schedule I have seen in this business was set by the supplier’s capacity: the agency’s bench, the next free sprint, the developer who could start in March. Now the schedule is set by what the client needs and when they can get to it.
What I have instead of a triangle is one measurement: hours a day, and who they were for. Mine came back first. You cannot give anyone their time back while you have none. But the promise is the second half, and the second half has a number: twenty-nine long days in twelve weeks.
Next September, that is the line to check.