Gregory Kurtzer, the creator of CentOS and Rocky Linux, has developed a new operating system kernel using an AI coding agent, completing the project in just 11 days. The experimental kernel, called Ginux, contains approximately 139,000 lines of implementation and test code and can boot Rocky Linux 10 through systemd.
The project demonstrates how AI agents can perform complex software engineering tasks under the direction of an experienced developer. But Kurtzer’s research acknowledges that Ginux remains an early-stage prototype with significant limitations in security, compatibility and hardware support.
Kurtzer used OpenAI Codex, powered by GPT-5.6 Sol, to develop Ginux through a process he calls Human-Directed Agentic Engineering (HDAE). The AI agent performed the bulk of the programming work, producing 1,097 Git commits during the 11-day development period.
Rather than having the AI agent operate independently, Kurtzer established the technical architecture, development requirements and acceptance criteria. The agent then generated code, ran tests, identified problems and revised its work.
This division of labor is central to HDAE, which Kurtzer documented in a technical research paper. The approach places responsibility for engineering decisions with the human developer while delegating implementation and troubleshooting to AI.
Ginux Runs Linux Software Without the Linux Kernel
Unlike conventional Linux distributions, Ginux doesn’t use the Linux kernel. It is a separate OS built primarily with Rust, a programming language designed to support memory safety. The new kernel targets x86-64 processors and includes its own scheduling, memory management, networking, device management and filesystem capabilities.
To support existing Linux applications, Ginux implements the Linux application binary interface (ABI). This compatibility layer enables selected Linux programs to operate without modifying their underlying code.
One of the project’s major technical achievements is its ability to boot Rocky Linux 10 and reach systemd’s multi-user.target, a milestone indicating that the OS has started its multi-user services. Ginux also successfully executes selected Alpine Linux programs, offering another demonstration of its Linux compatibility layer.
The kernel boots directly through Unified Extensible Firmware Interface (UEFI), without relying on an underlying Linux kernel. It can also retrieve and verify Open Container Initiative (OCI) images and use them to construct filesystems for isolated execution environments.
These capabilities suggest that Ginux could eventually support a range of Linux applications. However, its current compatibility is limited, and the project hasn’t demonstrated the broad hardware support required for enterprise deployment.
Ginux doesn’t establish that AI can independently produce enterprise-ready systems. But it offers a noteworthy example of a major trend in software development: experienced engineers are using AI agents to complete substantial programming projects in far less time than traditional development methods would typically require.
The speed of Ginux’s development is notable given the complexity of OS engineering. Kernel development requires coordinating numerous low-level functions that manage how software interacts with hardware. Still, producing working code is only one part of building a reliable operating system.
Perhaps most important, the project highlights the continuing importance of human expertise in AI-assisted development. The agent did a lot of work, but Kurtzer’s background in operating systems enabled him to establish technical requirements and evaluate the agent’s progress.
