A kernel built in eleven days by one agent task, then set aside by its author

Kurtzer's report describes what he calls Human-Directed Agentic Engineering. The human defines architecture, constraints, milestones, acceptance criteria and stopping points. An agent then implements, runs, diagnoses and revises. The case study is Ginux. Per the report, over eleven days one persistent coding-agent task produced an independent Rust-first x86-64 kernel that boots through UEFI, implements its own core operating-system mechanisms and a Linux compatibility layer, runs selected unmodified Alpine Linux programs, and boots a Rocky Linux 10 userspace with systemd through multi-user.target.

The report is plain about limits: it calls Ginux a research prototype, with narrow compatibility and hardware coverage, no independent security audit, and no claim of production readiness. The repository README goes further. It says the standalone-kernel phase is complete and preserved on a branch and tag, and that development has moved to a userspace kernel that runs as an ordinary Linux process and decides, per syscall, whether to emulate, mediate or deny. Kurtzer's stated reason is that reimplementing kernel mechanisms does not reproduce Linux's stable interface and ecosystem. He also writes that "the prototype was evidence, not an obligation to turn every experiment into a product."

The README says each approved gate records its scope, non-goals, acceptance criteria, verification, evidence and stopping point in a hdae/ directory, and that compatibility claims apply only to named contracts and executable evidence. The repository showed 498 stars and 59 forks when we read it. The Hacker News submission had 29 points and 50 comments about three and a half hours after posting.

Why it matters: This is a first-party account of a long agent run governed by written gates and executable tests, from someone who then used the result to decide what not to build. If you run long agent tasks, the gate files are the part to read. We read the report's record page and the README, not the full PDF, and we did not run the tests, so the eleven-day figure and the boot results are Kurtzer's claims, not something we reproduced.