My wife and I wanted to play Doom: The Dark Ages together. It wouldn't run on our PS4. So I started pricing a PS5, and I wanted to know one thing first: could it connect to the Bluetooth speaker built into our headboard?
No. Sony only lets Sony-approved Bluetooth devices connect.
Why does Sony have a say in what connects to the Bluetooth on a machine I bought?
So I didn't buy one. I built a Steam machine instead. Bazzite, my own Bluetooth card, thirty-dollar controllers, RGB, the whole thing. It sits by the bedroom TV and works like a console. We've played Portal co-op on it. It connects to whatever I tell it to.
I tell that story because it's the same reflex behind every AI agent in my house.
The House
I run a house full of AI agents. One runs the server and my side projects. One is my personal wellbeing system. One works for my wife. One runs our media and fixes everything else. Each has its own job, its own database, and its own chat on my phone.
They haven't always come from the same company.
Today they run on Claude. Until recently, my wife's agent, Val, ran on OpenAI's Codex.
Why Two Companies
When we set up Val, my wife had one requirement: she liked ChatGPT's personality. So Val was a Codex agent. Fine. My job isn't to tell my wife which AI she likes. My job is to build a house where it doesn't matter.
So nothing shared on the box assumes which AI is on the other end. Every agent gets a plain Linux user, a plain MySQL database, a plain Telegram bot. No vendor SDK in the plumbing. No "works best with" anything. If an agent can read a file, run a command, and send a message, it can live here.
That decision got tested sooner than I expected.
The Evening Val Switched
Val wasn't working out. My wife wasn't happy with her personality.
So one evening we moved her. Same Linux user. Same home directory. Same Telegram bot, so the chat on my wife's phone didn't even blink. We swapped the engine underneath, and her whole conversation history was sitting right there in a plain file for the new Val to read.
That's it. One evening.
A few hours later I could already tell she was getting better responses than she had on Codex. Right now, for us, Anthropic is winning. Right now.
And Codex is still on the box. If Val ever wants her old personality back, it's the same evening in reverse. Nobody had to pick a team.
The Model Isn't the Product
Here's what actually carries the weight in an agent: the instructions, the memory, the tools, the data. All of that lives in plain files and a plain database that I own. The model reads it.
I ran Opus 4.6 in my work pipelines long after newer versions came out, because it was working fine. Newer models mostly made things slower, not better. When a model upgrade finally did make a real difference, switching was a config change. The pipeline didn't care. Neither did the agents at home.
If your agent's memory lives inside one company's product, the agent belongs to them. If it lives in your database, you can hand it to whatever model is best this month.
The Same Playbook
I've written about how the cloud was sold as freedom and delivered dependency, and how AI companies are running the same play. Build on our platform. Use our memory. Use our agent framework. Don't worry about what happens when you want to leave.
Leaving should be boring. For Val, it was. That's the test I'd hold every AI setup to, at home or at work: if a better model shows up tomorrow, how long does it take to switch? If the answer is "we'd have to rebuild," you don't own your agents.
Sony doesn't get a say in my speakers. No AI company gets a say in my house.