Linux kernel debates AGENTS.md to manage AI coding agents
Linux kernel developers are discussing whether to add an AGENTS.md file to the source tree. The file would set ground rules for AI coding tools that increasingly submit patches to the project. The conversation is live, unresolved, and pulling in opposing camps.
What AGENTS.md Actually Is
Think of it as a README written for machines. It sits in the root of a repository and instructs AI agents on coding style, review processes, licensing, and attribution requirements. Other projects already use the format. The linux-firmware.git tree added one in June 2026, co-authored in part by Anthropic’s Claude Opus 4.8, covering how agents should add firmware blobs while respecting licenses and using Assisted-by tags.
The Linux Foundation backs the format through its Agentic AI Foundation, formed late last year. Anthropic, OpenAI, and Block contributed the spec alongside tools like the Model Context Protocol. Linux Foundation executive director Jim Zemlin said the goal is transparency and stability through open governance. OpenAI technical staff member Nick Cooper added that shared protocols matter for trustworthy agent infrastructure.
The Attribution Fight
Longtime maintainer Sasha Levin proposed a broader RFC last year requiring AI-assisted patches to carry Co-developed-by or Assisted-by tags with model details. The goal was transparency. The result was friction.
Prominent kernel developer Christian Brauner pushed back hard. In a mailing list thread covered by Phoronix, he argued the tags turned git history into free advertising for proprietary AI companies. The signal-to-noise ratio was low, compliance was inconsistent, and the data collected seemed pointless. He proposed dropping the disclosure requirement entirely.
“I find it very irritating that our git history has effectively started to function a bit like a free advertising platform for a bunch of AI companies and their proprietary agents and models.”
Greg Kroah-Hartman, the kernel’s second-in-command, occupies a more nuanced position. He uses AI tools himself with reported success. He has also moved to reject LLM-generated patches in the kernel’s staging area, except for genuine security fixes, citing the risk of overwhelming human reviewers and diluting the newcomer-friendly purpose of that part of the tree.
What Other Projects Are Doing
The Apache Software Foundation published draft guidance on AGENTS.md files just days ago with strict rules: agents cannot publish directly, cannot approve their own work, and attribution must distinguish generated output from assisted output. The document treats agents as tools, not contributors, and requires copyright notices to stay attached.
LLVM held a similar debate. Nick Desaulniers opened that discussion with a minimal draft file pointing agents to existing documentation. The project’s multiple subprojects with differing norms complicated any single-file approach, a challenge the Linux kernel faces at far greater scale.
Debian developers voted this summer on LLM usage, with options ranging from outright bans to cautious acceptance.
The Core Problem No File Can Solve
Critics of the AGENTS.md approach make a fair point. Kernel decisions hinge on subtle trade-offs between performance, security, and long-term maintainability. A markdown file can spell out process rules. It cannot encode judgment. An agent that reads the file might follow the letter while missing the spirit, and the existence of the file could invite more automated patch submissions that cite it as justification.
Proponents counter that clear instructions could steer agents toward useful work and reduce noise by defining what the project actually wants, covering everything from respecting subsystem maintainers to avoiding change types that humans handle better. They point to the linux-firmware repository as evidence the format works without chaos.
Kernel subsystems vary enough that any single file would need to acknowledge the hierarchy or risk being ignored. Proposals for the spec on its GitHub repository include progressive disclosure, where agents read deeper context files as needed, as one potential solution.
The outcome of this debate will likely influence how major open source projects handle AI contributions well beyond Linux. For now, the discussion continues without a decision.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
