NederlandsRocky Linux creator builds Linux-compatible kernel in eleven days with an AI agent
Gregory Kurtzer, known for CentOS and Rocky Linux, built Ginux in eleven days: a new core for an operating system that can run existing Linux programs. An OpenAI AI agent wrote almost all the code. It is still a research prototype.
Gregory Kurtzer, the creator of Rocky Linux and founder of CentOS, has built a new operating system kernel in eleven days that can run Linux programs. The project is called Ginux. It is not a modified version of Linux, but a kernel of its own, written in the Rust programming language. Most of the work was done by an AI agent: one continuous task in OpenAI Codex with the model GPT-5.6 Sol.
A kernel is the heart of an operating system. It shares memory and computing power between programs and handles contact with the hardware. Building one is considered among the hardest jobs in software.
Kurtzer is a well-known name in the Linux world. He created CentOS, a free version of Red Hat Enterprise Linux. When Red Hat changed course with CentOS, he started Rocky Linux as a community alternative. He is also a co-founder of CIQ, the company that provides support around Rocky Linux.
What Ginux can and cannot do yet
Ginux has its own memory management, its own way of letting programs take turns using the processor, its own file systems, its own networking and its own drivers, all written from scratch. Even so, existing Linux programs can run on it unchanged. This is because Ginux follows the same rules that programs use to talk to the Linux kernel. For Ginux, Linux is a kind of "face" that it shows to programs, not the core underneath.
The first results are impressive. Ginux starts Rocky Linux 10 up to the point where the system is ready for users. Several programs from Alpine Linux, a small version of Linux, also work. Ginux can also run programs in strictly separated compartments, each with its own processes, memory and files. According to the project, these boundaries are stricter than those of the containers that are common in data centres today.
It is a research prototype, however. For now Ginux only runs in a virtual machine, an imitation computer inside another computer. Real hardware, graphics cards and fast NVMe storage are not yet supported, and nobody has checked the security independently. Linux 6.12 has 375 system calls: the fixed requests that a program can make to the kernel. In Ginux, 82 of them work fully and 245 partly. Ginux deliberately refuses the other 48. Kurtzer himself states clearly that Ginux is not meant to replace Linux or to be used in businesses. The code is public, under the Apache 2.0 licence.
How the collaboration with the AI went
Kurtzer calls his method Human-Directed Agentic Engineering: building with an AI agent that is steered by humans. The human decides the structure, the limits, the interim goals and when something is finished. The agent does the rest: writing code, starting up, testing, looking for errors, improving and starting again. In eleven days this produced about 139,000 lines of code and tests, spread over more than 1,000 saved changes. By his own estimate the experiment cost about 8,000 dollars (about 6,900 euros).
Kurtzer deliberately wrote no code himself. He even closed his programming environment to resist the temptation, and only steered by making his requirements sharper. He had to step in once, when the agent made little progress for two days on a task that had been described too broadly. According to Kurtzer, Codex stayed on target better than the advanced Claude models from Anthropic, which more often wandered off onto interesting side paths.
He chose Rust partly because the language already stops memory errors before a program runs. In his view, programming agents also write Rust more reliably than C, the language in which Linux is largely written. The G in Ginux officially stands for GPT, with a nod to Kurtzer's own first name.
What does this mean for the road to the singularity?
This experiment shows how far AI agents have come in large, independent programming work. One experienced developer who only gave direction got a working kernel in eleven days that starts a full version of Linux. Such system software has so far needed many more people and much more time. Kurtzer himself names where the limit lies now: having code written is fast, but security, support for real hardware and years of maintenance remain the real work. The next step is for such agents to take over that part better and better too, so that ideas for new software can be tested much faster.
What is the technological singularity? Read the explanation →
