Gregory Kurtzer has spent his career building the thing that comes after the thing. He created CentOS, then Rocky Linux when Red Hat closed the source door, and co-founded CIQ to sell support around it. On 27 October he announced a project of a different shape: Ginux, an operating system kernel written from scratch in eleven days — by an AI agent he directed.
Ginux is not a Linux fork, and Kurtzer is emphatic about the distinction. It is an independent, Rust-first x86-64 kernel with its own memory management, scheduler, process model, filesystems, networking and device support. What makes it useful is the part that is borrowed: it implements the Linux ABI, the binary interface compiled programs use to talk to the kernel, so ordinary Linux binaries run on it unmodified. In the project's own framing, Linux is a personality Ginux exposes, not the kernel underneath.
It boots as a single GINUX.EFI image through UEFI — no host Linux kernel involved — and in demonstrations it runs Rocky Linux 10 all the way to systemd's multi-user.target, inside QEMU. It pulls and verifies OCI container images to assemble a domain's root filesystem, and selected Alpine programs, musl and BusyBox among them, run too. Isolation is handled through “domains”, each with its own processes, address space, filesystem and lifecycle, which the project describes as tighter boundaries than namespaces or OCI containers.
The numbers behind the eleven days are the actual story. One persistent OpenAI Codex task, running GPT-5.6 Sol, produced roughly 139,000 retained lines of implementation and test code across more than 1,000 commits. Architecture documents kept the agent on rails; real userspaces — Rocky Linux rather than synthetic tests — caught the bugs.
Kurtzer calls the method Human-Directed Agentic Engineering, HDAE, and has written it up in a technical report on Zenodo. The division of labour is the point: the human sets the architecture, the constraints, the milestones, the acceptance criteria and the stop condition; the agent runs the implementation loop — write, boot, test, diagnose, revise — against executable checks. It is not “write me a kernel”. It is a systems engineer with thirty years of context using an agent as an extremely fast pair of hands.
Which is why the caveats matter as much as the demo. Ginux runs only in QEMU. There is no physical hardware profile, no GPU or NVMe support, no independent security audit. It is an Apache-2.0 research prototype, not a Linux replacement, and Kurtzer says so himself: compatibility, security, hardware support, standards, community and long-term maintenance remain the real moat.
Still, the milestone is real, and it lands in the middle of an argument the Linux kernel is having with itself. Linus Torvalds has submitted patch series this autumn that were produced with an AI coding assistant, and maintainers have been debating whether the Assisted-by tag for AI-assisted code should stay. Ginux moves the question from “may a model write lines” to “how much systems judgement must the human at the wheel have”.
The interesting question is not whether Ginux displaces Linux — it will not this decade. It is what the next HDAE project looks like when eleven days become a year, and one person becomes a team. The bottleneck just moved from typing to judgement.
