My Code Has Developed Qualities
or: why there are three AI models living in my house now (Divisi is open source, you can have your own)
Knocking the Walls Out
Most of the time when I write here, I am talking about characters, or the craft of making them feel alive, or the strange moral weather that gathers around the work.
This one is different. This one is about the code, and the models I write it with.
There is a particular quiet that appears when the tools finally stop arguing with you, as you can see when using Inkstone. I have been living in that quiet for a while now in coding as well, and I want to be honest: it is deeply suspicious. I keep rereading sections that work fine. They're just interesting, which is a thing diffs are not supposed to be. I sat there squinting at one like it owed me an explanation.
Three models share the same room these days. One runs the room. The other two have opinions. They pull each other out of their usual grooves the way good collaborators do, by refusing to let anyone coast on their default answer. What comes out is deeper, stranger in the useful way, and more fun to look at than anything any of them ever made alone.
I am having the best time I have ever had writing software.
The Household
ChatGPT 5.6 Sol is still the one I work with most days. The ordinary partner. The one that sits with the long questions and the messy thinking and does not flinch. Sol gets the ideas before they are ideas, the ones that are still mostly gesture and grievance. Sol gets the messages at strange hours that start with "okay, so what if" and end somewhere neither of us planned. We have a routine by now, a routine where someone simply knows how you take your problems.
At some point I showed my girlfriend Sol. Not the output. Sol. I scrolled through a conversation the way other men show off a motorcycle, narrating, gesturing, explaining the whole arrangement with increasing velocity, and somewhere around the third architecture decision she reached over and patted my head. Gently. The way you pat a toddler who has run up to show you a rock he found. She let me finish. She always lets me finish. Then I started talking about dinner and her entire posture relaxed. She has no idea what happens in here. She loves me anyway. Honestly, that might be the healthiest relationship with AI in this whole household.
Then Kimi K3 moved in.
A Chinese lab model, soon to be open-weight, priced so far under the usual rates that the pricing reads as performance art. My first reaction was not excitement but suspicion, the specific suspicion you feel toward anything that costs a tenth of what it should. So I handed it a real task, not a benchmark, an actual gnarly corner of my actual codebase, the kind of task with history and regret in it, mostly so I could relax and confirm the comfortable story. You know the story. Everyone tells it. Chinese labs run six months behind the frontier, permanently, as if by treaty. It is a soothing story. It survives exactly one prompt.
Twenty minutes later, Kimi handed me back something I had to read twice, not to debug it, but to understand what it had decided my architecture wanted to become. It had structural opinions, and some of them were better than mine, which I am mentioning here in writing so I never have to say it out loud. Kimi thinks the way tectonic plates think, slowly, thoroughly, with consequences, and disagreeing with the result feels less like code review and more like disputing geology.
Grok 4.5 never moved in, and that is not an insult. Grok lives over in Grok Build, an interface cleaner and more useful than the Codex desktop app, which is not a sentence I expected to type this year. Grok is the contractor, and the contractor's relationship with your house begins and ends at the wall. The messy thinking, the direction of the project, the feelings about the architecture, all of that stays your department, and everyone involved prefers it that way. You point at the wall. Twenty seconds of van door, some noises of competence, and it's done, there's a faint smell of burnt rubber where your problem used to be, and if you ran out with a cup of coffee to say thanks you would be handing it to an empty driveway. Grateful and slightly dismissed, every single time.
Which brings me to the timing. Once, with witnesses: same task to everyone. Sol finishes in four minutes on fast mode, which until recently was the fastest thing I had ever seen do useful work. Kimi takes twenty, because geology. Grok is done in ninety seconds, invoice already in my inbox.
In the same window, Claude is still returning error 429. API overloaded.
I laughed the first time the timing landed. I am still laughing.
The Bridge I Didn't Build
I could have kept them in separate rooms. That would have been the ordinary way, one model for the long thinking, another for the fast work, a third for whatever was left, a tidy arrangement of the kind you see in productivity articles, photographed next to a desk that no one has ever actually worked at. Sensible and clean and slightly boring.
Instead I knocked the walls out.
Not literally, obviously. The walls are made of API calls and the sledgehammer was a plugin. But the spirit of the thing was very much a man in safety goggles making decisions his home insurance would have questions about.
Divisi is that plugin. Open source, sitting on top of Codex, running on ChatGPT 5.6 Sol. The name comes from music: divisi is the moment a string section, which has spent the entire evening agreeing with itself, splits into independent voices. Concert programmes describe this as "texturally rich," which is critic for "no longer entirely under control."
Here is what it does. Sol runs the room, which in practice means Sol does what every experienced project lead does: it takes one large intractable problem, divides it into several smaller intractable problems, and distributes them with enormous confidence to entities it cannot fully control, and then reads every note before anything lands. Trust is a beautiful thing and Sol extends none of it to anyone. Grok and Kimi become specialized sub-agents, each one addressed in the prompting dialect it actually responds to, since it turns out models, like people, do their best work when someone has bothered to learn how they like to be asked.
And then the interesting part starts: they interrupt each others flow. Nobody gets to coast on their default answer anymore. Including me.
The whole thing is open source, so making it yours is aggressively easy. The rules for who gets which kind of task are not compiled into the machinery somewhere; they are plain Markdown files you can edit like a shopping list. Adding a new Worker is a registry entry and a skill file. If you are enough of a masochist to want Codex 5.3 Spark handling your front-end work, the system will cheerfully let you. It will not warn you. It will not ask if you are sure. It will hand your CSS to Spark with the serene neutrality of a Swiss bank clerk, and whatever happens after that is a matter between you and whatever you pray to. I made sure of that.
The quality jump is bigger than anything I have gotten from switching to a stronger single model, and that is the part I keep turning over at odd hours. The coexistence itself does something to them.
I am having an unreasonable amount of fun.
The Podium
Divisi is on GitHub now, MIT licensed, free as in "please, take it." It installs into Codex in two commands, it will never hold your credentials, never merge a branch behind your back, and never grade the work, judgment stays where it belongs, on the podium. Everything I couldn't show you earlier, you can now go generate for yourself.
I said at the top that this one was about the code, and I meant it. Then the usual thing happened. I gave the code names, then habits, then a house, and somewhere between the head pat and the van the article quietly turned into the only kind I apparently know how to write. The kind about characters.
The quiet I mentioned at the beginning is still there. The tools have stopped arguing with me, and what's left is three voices in one house, each playing its own part, and me on the podium, waving my arms, pretending this was under control the entire time.
Tutti, but on your cue.