docs: try to reduce amount of AI slop

[gudvinr]

Have you followed these guidelines?

• I have tested my changes
• There are no breaking changes
• I have thoroughly tested my changes and verified there are no regressions
• My commit messages follow the Conventional Commits specification
• I have read and understood the contribution guidelines

---

This is the adaptation of AGENTS.md file from llama.cpp/AGENTS.md and OctoPrint/AGENTS.md.

Idea here is to prevent contributors from mindlessly dumping agents output

• There's no one who can be responsible for security issues
• These PRs usually extremely huge and unstructured. This increases strain on reviewers (that's basically one person here)

Influx of fully generated PRs where author has no meaninful input proven to increase pressure on reviewers.
See Godot developers experience:

• https://www.gamedeveloper.com/programming/godot-co-founder-says-ai-slop-pull-requests-have-become-overwhelming
• https://www.devclass.com/ai-ml/2026/02/19/github-itself-to-blame-for-ai-slop-prs-say-devs/4091420

There's no shame in using an assistant who can do tedious tasks. llama.cpp has this kind of disclaimer and you can't say that llama has no beef in this AI game.

Of course this won't help against determined individuals who violate any and all instructions. But at least that means they do that willingly.

[jvoisin]

I'm not thrilled about this, for a couple of reasons:

• We shouldn't care what tools people are using. If someone wants to use AI or whatever, it's their problem. We don't have rules for using an IDE or ed for example. We only judge on the quality of the PR.
• On a personal level, I loathe the fact that AI is a thing and that's it's becoming pervasive, despite being a multilevel strain (energy, materials, labor, licenses violation, open web destruction, loss of cognitive capabilities, …). Any encouragement of this tendency makes me sad and angry at the current state of the world.
• Large pull-requests can (and should) be closed, whether they're AI-generated or not, "this isn't reviewable."
• Owing to the previous point, people are responsible for their pull-requests. If they're sending slop, they should simply be banned. Handhodling their agent isn't doing anyone any favours.
• People opening slop PR likely won't bother with the AGENT.md file, and/or tell their agent to simply ignore it.

[gudvinr]

We shouldn't care what tools people are using.

I clarified that it's not about forbidding usage of AI as a tool:

There's no shame in using an assistant who can do tedious tasks.

That's purely to deter contributions containing completely generated code. However, as you noted:

People opening slop PR likely won't bother with the AGENT.md file, and/or tell their agent to simply ignore it.
If they're sending slop, they should simply be banned.

These are people who know what they do is against the rules. This AGENTS.md does say that "Repeated sending of AI generated contributions can even lead to a ban". So if they want FAFO then well, that's their decision.

I cannot understand motivation behind this but not all people like that. Usually, when people being told that something is not wanted or wrong they stop doing that.

Large pull-requests can (and should) be closed, whether they're AI-generated or not

Sure thing, but still takes non-zero amount of time. If this PR won't exist in the first place, it wouldn't need to be looked at. That's basically the only reason for this thing to exist.

---

See llama.cpp#18388 for more background behind this. That's the opposite of encouragement in my opinion.

[fguillot]

AI-generated pull requests are not necessarily forbidden. Contributions that do not meet the project guidelines will not be merged, regardless of whether the code was written by a human or an LLM.

For example, using AI for a small, well-understood change that can be fully reviewed by a human is fine.

LLMs are improving quickly and are becoming part of many developers’ toolboxes.