A System Above the Vendor: Independence as a Principle

Not long ago Claude Code went down. Second time in one day - server problems on their side.

Usually I just wait and do other things. But that time I finally did something from my “someday” list: for the first time I switched my brand agent to Codex from OpenAI. The switch took one second.

This article is about that second. Because behind it stands the principle my whole AI system is built on: the system lives above any vendor, and none of them can take it away with them.

The knowledge lives in my files, not at the vendor’s

My whole system - the agents, their rules, the accumulated knowledge, my voice, the memory of my projects - lives in plain markdown files. Not inside Claude, not inside Codex. On top of them. The model is an executor that comes to work with my files. When one tool goes down, the work does not stop: I take another one and continue from the same place.

I built the system from the start to survive any of them going down. But building is one thing, checking is another: it was scary to just switch and see what happens. It turned out to work exactly as planned.

The other side of this principle is closed boxes. A Custom GPT, for example: you can’t add anything to it from outside, and, more importantly, you can’t take anything out of it. Everything you put in - wordings, rules, everything you worked out - stays with the vendor. I already wrote in the second brain article about how open files let agents improve each other - here one thing matters: all of it is possible only because the knowledge sits in a simple, open form. With me, not with them.

This is what freedom without losing safety means. Not “don’t use the big models,” but “use any of them - in a way where losing any of them takes nothing away from you.”

Changing the tool is changing the lens

Then something happened that I didn’t plan.

I asked Codex what it knows about me, asked it to look at what we’ve been building here and give its opinion - from the outside, with a fresh eye. It gave me nine points. Some - exactly right, I took them into work right away. Some - I see differently. But the real lesson was not in the points.

Changing the tool worked like changing a lens. Codex saw what had gone blurry for me: where I over-complicated, where old and forgotten things stayed. Your own agent gets used to your work the same way you get used to it yourself. And an outside look - even a competitor’s look at a colleague’s work - opens up the familiar.

So the item “backup for when it goes down” also turned into “a second opinion on demand.”

Not one best, but a team

After that I didn’t go back to “one main model.” Now two models work inside one agent, and sometimes I feel like a matchmaker.

Claude understands me: it writes almost in my own voice, catches the tone, jokes in the right places. Codex draws and critiques. Its images come out noticeably cleaner, ready to go straight on the site. But it writes like it’s leaving a comment in code: dry, to the point, very much like an engineer. So there I am between them, trying to marry an artist and an engineer. To one: “write a warm text.” To the other: “you handle the images.” And then I watch so they don’t rewrite each other’s work.

The most interesting part: this doesn’t drain me, it makes me happy. I used to look for the one and only best AI. Turns out the point is not to pick one, but to build a team and give each the work it does best. Just like with people.

And this is possible for the same reason the switch took one second: the system is simple files on top of the models, so adding another “teammate” (another LLM) takes a couple of minutes. One writes, one draws, and I’m the director deciding who is on next.

“Which model is better” is a question that simply disappears with this architecture. Better - for what? Today one is stronger in texts, tomorrow another in images, the day after a third one comes out. While the system lives in your files, every new jump of any model is your win, not your migration.

The same principle, one level up

Recently I found the Liberman brothers and their project Gonka: a network where compute for AI is spread across thousands of independent participants, instead of being locked in the data centers of five companies.

What caught me is how they combine a worldview with an engineering observation - exactly what I see in my own work every day. I don’t build one all-powerful agent. I build systems of many small ones: each with a narrow task, its own character, its own zone of responsibility. And again and again I see the same thing: a swarm of simple, independent units turns out smarter and more alive than one huge center that tries to hold everything in one hand. When everything depends on a single brain, one error, one failure, one wrong order from above is enough, and all of it falls at once. That is a single point of failure - one lever someone will pull one day. Where there are many independent nodes, one drops and the rest pick it up, and the whole keeps living.

Distributed is noisy, ugly, unclear - and somehow it survives and outperforms. This is not about crypto and not about a trend. This is how complexity works in general, from neurons to living markets. And when the infrastructure for AI starts to be built on the same principle, everything lines up for me: yes, this is how it should be.

My system of markdown files on top of replaceable models is the same principle, only at the level of one person. Don’t tie yourself to a single center. Don’t put all the knowledge in one hand. Keep the nodes small, independent and replaceable.

Where your files live

The future of AI should not belong to five corporations. I know how loud that sounds, and I don’t like manifestos without practice. But this loud phrase has a very quiet and simple beginning: look at where your knowledge lives. If everything you have built with AI - your rules, your voice, your memory - lives inside someone’s closed box, then it is not your system. You are its.

And if it is plain files you can read, edit and move, then any AI - the newest and most powerful one - becomes just another member of your team. It comes and goes. The system stays.