The obsession, counted.

On 26 August, Lex Fridman published his second conversation with David Heinemeier Hansson, the creator of Ruby on Rails. It runs five hours and twenty-two minutes. I listened to most of it at my desk, at night, with an agent building a feature in the terminal next to the player, which is the only honest way I could have listened to it.

DHH spent twenty years handcrafting code and being loud about how much the craft mattered. In that episode he says the quiet part out loud: agents now write the code, and he is fine. Not because the craft died, but because it moved. “When I’m working in Ruby code and asking agents to make changes, I still sweat the details,” he says, and then admits the economic payoff of that sweat is shrinking fast. What is not shrinking is the part he calls taste. “Most organizations don’t know what they want. They’re bottlenecked on ideas, vision, taste.”

He draws a line I have been trying to draw for months. Vibe coding is telling an agent to build something and never looking at what it built. Agentic engineering is the same speed with the details still sweated, by someone who knows what wrong looks like. And he keeps coming back to one word that has nothing to do with software: mission. “I have a mission. I’m going somewhere. I can channel all this new power towards a singular goal.”

That is the piece most people miss when they talk about AI and building. The tools are real. DHH puts it as decades of progress arriving in nine months. But power without a direction is just a faster way to be busy. Obsession is the precursor. The tools only decide how far the obsession gets to travel.

You can find the episode at lexfridman.com. This post is what it made me want to do: count my own obsession honestly, and say where it came from.

Here is the part I left out of the first post, and DHH’s conversation is the reason to put it in now.

I did not come to this as a manager who discovered a magic box. I came to it as someone who has been writing code, on and off, for most of his life. In school, in the early 2000s, I won Olympiads in coding. I took an A in every computer course I ever sat, at school and again in engineering. I worked briefly at Infosys as a systems engineer, which is where I learned how open systems actually fit together, the unglamorous plumbing between one machine and the next. Over the years I have dabbled in SQL, C++, Java and, as it happens, Ruby on Rails, which is the framework DHH built and the reason I first knew his name.

None of that makes me a great programmer. It makes me a person who knows what wrong looks like. When an agent hands me a query that will scan a whole collection on every request, or a permission check that trusts the wrong field, or a migration that would wipe what it was meant to move, I recognise it, because I have written all three by hand and paid for them. That recognition is the entire difference between the two things DHH separates: vibe coding, where you never look, and agentic engineering, where the speed is the same and the details are still sweated.

DHH says on the episode that he hates the word agentic. Loved or hated, call it what you like. My position is that to do it today you must know the nuts and bolts. Maybe tomorrow you will not. The tools are moving fast enough that I would not bet against it. But that is not today, and the people shipping real systems to real academies at three in the morning are the ones who can read what the agent wrote.

Which brings me back to his three bottlenecks: ideas, vision, taste. The first two you can bring from anywhere. My ideas come from the hall, not from a terminal. The third one comes back to the nuts and bolts every time. Taste in software is the residue of having done it by hand, of knowing which of the three options the agent offers will still be standing in a year. And that is the hole we are heading into right now. If the next generation never writes the bad query themselves, where does the taste come from? I do not have the answer. I only know that mine came from twenty-odd years of doing it the slow way, and that the agent gets it for free every night.

The repository opens on 2 February 2026 with an empty commit from an AI agent. The next morning it receives a working prototype: 77 files, about 6,800 lines. That prototype was built on Emergent, an AI app builder, in the weeks before, and the weeks before that are not in git at all. So the first honest thing to say about the count is that it starts late.

Emergent kept auto-committing until 5 March. Fifteen active days, 1,165 commits, and a codebase of roughly 61,000 lines by the end of it. On 6 March I took the keyboard myself, with Claude Code, and the repository changed character overnight. Auto-commits became hand-authored, reviewed, tested work. From that day to this one there are 1,280 of my own commits across 94 active days, 97 pull requests, and about 153,000 net new lines.

The Emergent month is under-counted by construction. One auto-commit can hold a whole afternoon. Treat every Emergent number in this post as a floor.