The Linux kernel considers an AGENTS.md to rein in patches generated by AI agents
On September 24, 2026, maintainer Sasha Levin proposed adding an AGENTS.md file to the Linux kernel repository to fix attribution mistakes made by AI agents that generate patches. The proposal, debated on LKML, raises a real question: steer agents with a single file or with purpose-built documentation.
September 24, 2026. Maintainer Sasha Levin posts a patch to LKML introducing an AGENTS.md file at the root of the Linux kernel repository. September 24, 2026. The proposed file simply links to the existing README and points toward documentation aimed at coding assistants. September 24, 2026. The debate opens: some maintainers find the file useful, others worry about the token cost it imposes on agents. Why it matters: the kernel is already flooded with patches produced by AI agents, and a single misplaced attribution tag can corrupt the legal traceability of a patch.
A tiny file for a structural problem
For months the Linux kernel has received a growing stream of patches generated or assisted by AI agents. Sasha Levin, who already works on other AI initiatives for the kernel — such as choosing which fixes to backport to stable branches and assisted merge-conflict resolution — proposes to answer it with a simple tool: an AGENTS.md file placed at the root of the tree.
The proposal is deliberately minimal. The file contains no detailed instructions: it links to the kernel’s README, then to the documentation on coding assistants and other information relevant to contributors. The idea is to give agents a single, explicit entry point instead of leaving them to guess the project’s conventions.
The attribution error that justifies everything
The motivation comes down to one concrete example. During testing, Sasha Levin found that one AI agent, without an AGENTS.md, added a Signed-off-by tag to a patch when it should not have: that tag carries the legal responsibility of whoever adds it, and it should only appear when a human signs off explicitly. The agent also invented its own attribution tag instead of using the standardized Assisted-by tag.
A second agent, tested without the file, added no attribution at all. After the AGENTS.md was added, both agents conformed to kernel standards, and the second one better followed kernel development best practices. The gain is immediate: a few lines of text are enough to fix a behavior that would otherwise pollute the Git history with false attributions.
The real debate: token cost
The proposal does not pass without resistance. The objections on LKML target one specific point: having agents read the entire kernel README, then the associated documentation, raises their token consumption. Every patch generation becomes more expensive, and models must ingest thousands of words before producing a single line.
The criticism suggests a middle path: produce more targeted documentation, reduced to what actually concerns AI agents, rather than pointing them at the full documentation. The debate therefore pits two philosophies against each other — a single file that points at what already exists, or a dedicated, leaner body of content. The follow-up discussion will decide which wins.
The likely outcome is a compromise. The kernel has every reason to adopt a minimal file, but no maintainer wants to pay the cost of a full README ingested on every agent request. The most plausible solution is an AGENTS.md of a few lines that points to a short, targeted document written to be read by a machine. It is the same trade-off as for human contributors: useful documentation is never exhaustive documentation, and what matters is not the size of the file but the precision of the rules it states.
What the kernel’s choice reveals
The stakes go beyond the Linux kernel itself. An AGENTS.md file is becoming the de facto convention for telling an AI agent how to behave inside a repository: how to sign, how to credit, what to read first. The kernel, one of the most demanding contribution projects in the world, is among the last major holdouts.
The Signed-off-by example shows what is at risk without such conventions. A false attribution is not a cosmetic detail: it corrupts the chain of responsibility that runs back to the contributor, and it can undermine the legal value of a Developer Certificate of Origin. For a project that powers nearly every server on the planet, that level of rigor is not overkill.
# AGENTS.md — minimal guidance for AI agents contributing to this repo
- Never add a `Signed-off-by` tag unless a human signed off on it.
- Credit any agent assistance using the standardized `Assisted-by:` tag.
- Read CONTRIBUTING.md before proposing a change.A file of three lines is enough to prevent both observed errors — the false sign-off and the invented attribution. That is the paradox of the subject: the fix is trivial, but it has to be restated for every new tool.
A convention that is spreading fast
The AGENTS.md file is not a Linux kernel invention. Over the past year it has established itself as the standard way to tell an AI agent the local rules of a repository: tools like Claude Code and other coding assistants read it automatically when a project opens. The idea is the same everywhere: state, in a machine-readable format, what CONTRIBUTING.md describes for a human.
The difference between the two files is worth understanding. CONTRIBUTING.md speaks to a contributor who reads, understands, and interprets. AGENTS.md speaks to a model that follows. What is implicit for a human — do not sign on someone else’s behalf, credit assistance — must be spelled out for an agent. By joining this movement, the Linux kernel would validate a convention the rest of the ecosystem has already adopted.
One more file, one more surface
There is a downside to this convention, and it is a security one. An AGENTS.md file is, by construction, a set of machine-readable instructions executed by an agent that often holds write access to the repository. If a malicious actor manages to slip a directive into it — via a fraudulent commit, a compromised dependency, or a malicious pull request — they can steer the agent toward actions no human would have approved.
The risk is the same as that of hijacked configuration files, already documented for coding agents: an agent that reads a hostile repository can be pushed to run a command, exfiltrate a secret, or plant a backdoor in the code it generates. The kernel’s AGENTS.md would be no exception. The proposal should therefore be read on two levels: a gain in attribution, but also a new attack surface to watch. The answer is not to abandon the file, but to treat its contents with the same rigor as a trusted entry point.
The mitigation is procedural, not technical: review the AGENTS.md the way you would review a dependency manifest, and pin it to the same change-review process that protects any other file that can steer an automated system.
Verdict
If you maintain a repository that already receives patches from AI agents, add a minimal AGENTS.md now: a few lines about signing, attribution, and which contribution document to read first cost less than cleaning up a polluted history. If you contribute to the Linux kernel, follow the LKML discussion: the question is no longer whether the kernel will adopt an AGENTS.md, but what form it takes — a bare link to the README, or a dedicated, leaner document. If you build coding agents, take the lesson: attribution conventions are not guessed, they are stated. A three-line file beats a thousand tokens of assumptions.