Technical thesis · October 2026

Towards World Models of the Operating System

AI agents decide what to do while they run, consulting a language model at each step, so their behaviour can no longer be read off their code. Every action they take still passes through the operating system kernel. We record agents at the kernel and find they take on distinct shapes that cross process boundaries and change from run to run. We then propose a world model of the operating system, a learned representation of how a machine's state evolves, as the foundation for detecting agents, reconstructing incidents and governing what any program on a machine can do.

Read as one piece or in three parts →