Story No. 04
Have we been replaced yet, Dario?
In January, Anthropic's CEO said models might soon do everything software engineers do. Nine months later, AI writes most of my code, and my work has multiplied.
It’s Sunday, and I’m at the gym. Between sets I’m not resting. I’m in a Telegram chat with an AI agent, answering its questions about the post you’re reading now.
What Dario said
In March 2025, Dario Amodei, the CEO of Anthropic, said that within three to six months AI would be writing 90% of the code, and within a year essentially all of it.
In January 2026, at Davos, he went further: “I have engineers within Anthropic who say I don’t write any code anymore. I just let the model write the code, I edit it, I do the things around it.” He gave it six to twelve months before models do “most, maybe all” of what software engineers do, end to end.
That window runs from July 2026 to January 2027, so we’re in the middle of it. Here is how it looks from a studio of one.
He was right about the code
I’ve been writing code for twenty years. On July 17 I made the first commit to the repository behind the Neon Labs games and apps. A hundred days later it has 255 commits, and 160 of them name Claude as a co-author. In September it was three in every four. Stefan, our sticky-note board, went from its first commit to its own address in five days.
I barely type code anymore. On that part, Dario was right.
He was wrong about the work
What I heard in those quotes was: less for me to do. The opposite happened. Writing the code got cheap, and “the things around it” grew to fill every free hour I had.
Some of those things were always there. My checklist for releasing Neon Comet opens with two sentences: “The code is ready. What’s left is clicking through the outside services.” Play Console, App Store Connect, RevenueCat, Supabase, Cloudflare, Vercel, Expo’s build service, Microsoft’s Partner Center. My instructions for clicking through them run to 4,366 lines, and no model clicks them for me.
The rest is new. A fast agent means one isn’t enough, so I run several. I look after eight projects now, some of them for clients. In October I made 485 commits across them, and this past week I committed to six. Some of the work runs in Claude Code on my computer, which I steer from my phone with Remote Control. Some runs through Hermes, an agent I talk to on Telegram, with a group for each project. Four scheduled jobs message me on their own.
Each of them stops to ask which option I want, whether this is right, whether it can push. An unanswered question is an agent doing nothing, so I answer.
In fairness, Dario said this part too. Back in 2025 he added that the programmer still has to say what the app is, what the design is, and whether it’s safe. I didn’t expect that sentence to be the whole job, eight times over.
Between sets
For twenty years, work stayed on the computer. I got up from the desk and it stayed there. A few months ago that stopped, and now the desk is in my pocket.
My family sees it. The phone is always with me, I check every notification, and I never make the AI wait.
I chose this, and I like it. But I’m on call for my own agents, and I’m the only one on the schedule. I haven’t found a way to switch it off, and I won’t pretend this post ends with one.
This post is the proof. It came out of that Telegram chat: an agent asked me questions, I answered them between sets, and it wrote the first draft from my answers.
So no, Dario, I haven’t been replaced. I just never leave work anymore.
— Kamil