Codex builds a Linux-compatible kernel in 11 days that boots Rocky Linux 10

What this is about

On October 27, 2026, Gregory Kurtzer announced on X that he had built a new Linux-compatible kernel in just 11 days. Kurtzer is a co-founder of CentOS and the initiator of Rocky Linux, which emerged after Red Hat ended CentOS Linux. The project is called Ginux. It is an independent x86-64 kernel written mostly in Rust, not a fork of the Linux kernel. It boots Rocky Linux 10 and runs systemd. According to the report, it comprises about 139,000 lines of code. The code is on GitHub under Apache 2.0, along with a 13-page technical report.

How Ginux came about

The model was an experiment by Anthropic's Nicholas Carlini: 16 Claude agents running in parallel built a C compiler, with about 2,000 sessions, roughly $20,000 and around 100,000 lines of Rust. A kernel is harder to verify than a compiler with a test suite. Kurtzer wanted to know how far AI can take systems software when a domain expert sets the architecture, constraints, milestones and acceptance criteria.

The work ran almost entirely in a single, continuous Codex task with GPT-5.6 Sol. The author defined the architecture and workflows, interpreted results and decided what to accept. Kurtzer calls his approach "Human-Directed Agentic Engineering" (HDAE), with two loops:

• Outer loop: Vision → architecture and constraints → acceptance criteria → human review of observed behavior → correction, prioritization or sign-off.
• Inner agent loop: Implement → build and run in QEMU (an emulator for virtual machines) → measure and analyze failures → revise → regression tests.

He chose systemd as the litmus test because it needs many system components and carries the entire boot process. Since it is tricky to start inside containers, Ginux built "domains" first and containers only afterward. On X, Kurtzer said Codex wastes fewer tokens on side paths. Even so, he often had to pull it back on track. He even closed his IDE so he wouldn't step in himself and would only give the AI new instructions. His recipe: lots of acceptance hurdles. "Vibe coding" without guidance doesn't suit a project like this. Many errors produced plausible-looking output that violated Linux semantics, isolation or lifecycle rules, in other words "AI slop".

What Ginux can do

The result: 1,037 commits, 261 tracked files, 63 Rust source files. Ginux has its own memory management, scheduler, process model, file systems, networking and device support. Because it implements the Linux ABI (the binary interface programs use to talk to the kernel), unmodified Linux binaries run on it. The kernel is launched directly via UEFI as GINUX.EFI, with no Linux host. It runs isolated "domains" with their own processes, address spaces, file systems and lifecycles. According to the project, this isolation is stronger than with OCI containers or Linux namespaces. Ginux can load and verify OCI images and build a domain's root file system from them. That lets Rocky Linux 10 boot up to the systemd target multi-user.target. Parts of Alpine (musl, BusyBox) run as well.

Benchmarks

The comparison used a VM on an Apple silicon Mac with 1 CPU core and 2 GiB of RAM. With Ginux, Rocky Linux 10 was ready to log in after 12.9 seconds. With the Linux kernel it took 42.6 seconds (about 3.3 times faster). That comparison is only partly fair, since Ginux doesn't have to load a kernel and the file system was cached. The Sysbench runs (3 seconds, median of nine runs) are more telling. The table with the numbers is missing from the source text. According to the report, CPU and threads come in slightly above Linux and memory slightly below, with deviations from single-digit percentages up to roughly ten percent.

Cost and license

API costs were about $8,000. The license is Apache 2.0.

Perspective

Kurtzer himself warns against reading too much into this. AI can produce complex systems software, he says, but compatibility, security, hardware support, standards, community and long-term maintenance are the real moats. Ginux has no comparable hardware matrix, no standards bodies and no community, and it can't promise that today's programs will run tomorrow. How robust and secure the kernel is beyond the tests remains an open question. The numbers come from the author and haven't been independently verified. The result is interesting for anyone who wants to know how far agentic coding carries on systems software: it shows a lot when an experienced expert steers, and little without one.