The Linux kernel has long been a favored target for experiments in automated code generation. Patches signed by humans, patches signed by pseudonyms and, increasingly, patches produced by LLM agents land in maintainers' mailing lists. Today, however, a different signal arrives: a patch set proposes adding an AGENTS.md file to the repository, a Markdown document designed to give instructions to LLM agents. Until now the kernel did not have one, even as the volume of AI-generated contributions has grown markedly.

The difference from a simple anti-spam filter is the direction of control. Instead of merely discarding low-quality patches, kernel maintainers would begin defining the rules an agent should follow to prepare and propose its contribution. An AGENTS.md file does not force anyone to comply, but it creates a verifiable reference point: if an agent ignores the file, maintainers can close the patch with less debate; if it follows the file, the mechanical part of the work becomes more predictable. It is a shift from reactive moderation to a behavioral specification.

This move has implications beyond kernel bureaucracy. Linux is a project that survives on a delicate balance of trust, reputation and control. When LLM agents produce code, the boundary between author and tool blurs. An AGENTS.md file does not solve the attribution problem, but it introduces a minimal contract for those using models to contribute to critical infrastructure. Teams working in self-hosted environments, where code generated by local models never leaves their servers, can see this proposal as a precedent: development policies should be made explicit for internal automated systems too, not only for human collaborators.

For kernel maintainers, the immediate benefit is a stronger basis for rejecting poorly formed automated contributions. For developers using LLMs in good faith, the file lowers the cost of understanding how to participate. For those producing low-effort patches, the existence of explicit instructions makes it harder to hide behind ambiguity. It is no coincidence that the proposal arrives at a time when the quality of LLM-generated patches is under scrutiny in many open-source projects.

The question is no longer whether LLM agents will participate in kernel development, but under what rules they will be required to do so. For those watching on-prem and local architectures, it is a reminder that governance concerns not only the models themselves, but also the artifacts those models produce.