Agentic OS explores a simple question: how does a computer system change when an AI agent can go beyond running individual commands to configure, monitor and develop the environment itself?
A small, declarative foundation
NixOS provides the foundation because the system can be described as a configuration and reproduced. Wayland supplies the graphical environment, while tmux preserves working contexts across longer sessions. The installation starts small by design; additional tools are added only when a specific task calls for them.
This makes it clear what the agent has set up. Instead of an accumulation of ad hoc manual changes, the system remains in a state that can be understood and traced.
The agent as a system operator
In this experiment, Codex works with administrative privileges. The agent can install software, adjust configurations, set up services and integrate tools it develops itself. What interests me is the whole cycle of observing, planning, changing and checking, rather than any individual command.
These extensive permissions are also the sensitive part. Mistakes can affect the whole system, not just a project folder. Reproducible configurations, clear boundaries and verifiable changes therefore matter even more than they do for an ordinary assistant.
What the experiment aims to explore
Agentic OS is an ongoing experiment rather than a finished product. It explores how much of its environment an agent should shape for itself and which decisions must remain human. I'm particularly interested in transparency, reproducibility and the boundary between useful autonomy and unnecessary risk.