AI Writes 139K-Line Linux-Compatible Kernel in 11 Days: Meet Ginux

A New Kernel Born from AI

Gregory Kurtzer, founder of CentOS and Rocky Linux, has built a brand-new operating system kernel called Ginux in just 11 days using an AI agent. Unlike Linux forks, Ginux is a clean-slate, Rust-based x86-64 kernel with its own memory management, scheduler, process model, filesystem, network stack, and device drivers.

Linux ABI Compatibility

Despite being an independent kernel, Ginux implements the full Linux Application Binary Interface (ABI). This means existing Linux executables run unmodified. As the project puts it, "Linux is a personality trait that Ginux exposes, not the underlying kernel." The name Ginux reflects its AI origin: the "G" stands for GPT, specifically GPT-5.6 Sol , which wrote the entire codebase via OpenAI Codex.

What Ginux Can Do Today

Ginux compiles to a single GINUX.EFI image that boots directly via UEFI without a Linux host. It runs isolated "domains" — each with separate processes, address spaces, filesystems, and lifecycles — providing stricter isolation than OCI containers or Linux namespaces. It can pull and verify OCI container images to assemble minimal root filesystems for domains. Using this mechanism, Ginux successfully boots Rocky Linux 10 all the way to systemd's multi-user.target. Some Alpine Linux programs (musl, BusyBox) also run.

Ginux is a research prototype released under Apache 2.0. It runs only in QEMU, lacks physical hardware profiles, GPU and NVMe support, and has not undergone independent security audit. It is not production-ready and cannot replace Linux.

Human-Directed Agentic Engineering (HDAE)

The core story is the development methodology Kurtzer calls Human-Directed Agentic Engineering . He published a detailed technical report on Zenodo. The division of labor is strict:

Human decides — defines architecture, constraints, milestones, acceptance criteria, and when to stop.

AI agent executes — writes code, runs tests, diagnoses failures, and iterates until executable checks pass.

For Ginux, this meant a persistent OpenAI Codex task running GPT-5.6 Sol for 11 days, producing ~139,000 lines of retained code (implementation and tests) across more than 1,000 commits. Architecture documents kept the agent on track, while real user-space workloads like Rocky Linux caught bugs that synthetic tests missed.

"Human direction" is critical. This is not a chat prompt like "write me a kernel, no bugs." It is a domain expert with decades of systems experience steering an agent through the heavy coding lift.

Expert + AI: How Far Can They Go?

Kurtzer cautions against overhyping the result: the real moat remains compatibility, security, hardware support, standardization, community, and long-term maintenance — things an 11-day agent run cannot replicate. Yet the achievement is striking: a from-scratch kernel that boots an enterprise distro with systemd, built in 11 days with GPT-5.6 (not even the latest 6.1).

AI will not replace experts, but experts who understand kernel internals can now compress years of manual labor into days. When deep domain knowledge pairs with frontier AI agents, projects that once required large teams over years start to look like weekend side projects. Ginux will not replace Linux, but it demonstrates how far and fast a skilled developer can now go.