Let’s Build an Agentic OS

Why the agent ecosystem needs a kernel, not another framework — and what I learned building one the hard way.

Nuno Roberto | Chief AI and Data Architect

There’s a scene that keeps replaying in my head. It’s 1972, and a programmer is sitting at a PDP-11 terminal, typing instructions directly into bare metal. No process scheduler. No memory protection. No file system abstraction. Just raw machine code hitting raw hardware, and if another program happens to be running — well, it isn’t, because there’s no mechanism to let two programs coexist without destroying each other’s state. The programmer doesn’t think of this as primitive. This is just how computers work. You write a program, you load it, you run it, and you pray that nothing corrupts the memory addresses you’re depending on.

Fifty-four years later, I’m watching the AI agent ecosystem do exactly the same thing.

I’ve spent the last two years building an agentic AI platform called Kairos for enterprise logistics, and somewhere around month fourteen, I had a realization that stopped me cold. Every agent I’d built, every multi-agent workflow I’d orchestrated, every tool integration I’d wired up, all of it was running bare metal. No process isolation between agents. No shared memory architecture. No resource scheduling. No access control model. No system call interface. I…