Re: AGENTS.md

> Date: Thu, 06 Aug 2026 09:03:48 +0300
> From: Jean Louis <[email protected]>
> Cc: [email protected], [email protected], [email protected],
>  [email protected]
> 
> On 2026-08-06 07:42, Eli Zaretskii wrote:
> 
> >> Did you test all the millions of LLMs on it?
> > 
> > Is this a serious question?
> 
> You said you tested LLM. Yet there are millions of LLMs. So I am 
> pointing out that it is better you express yourself specifically.
> 
> You probably tested just the software agent. LLM is normally not trained 
> on AGENTS.md so it is not relevant testing the "LLM".

Sean described what he did.

Once again, it is well known that not all LLMs automatically read a
file by this name.  Having this file is a partial measure, not a 110%
reliable method of preventing LLM-generated code in our repository.
This has been explained many times already, and yet you still keep
beating this long-dead horse.  The only effect of doing this is to
raise the noise level on this list.  Please stop.

> And yet in that testing, why not say in transparent what exactly did you 
> test?
> 
> - which software?
> - which LLM was tested?

If Sean wants to describe this, he will.  But it is immaterial,
because for our purposes it is enough to know that this measure is
sometimes effective.  Of course, if someone suggests changes which
will make it more effective with more LLMs, we will consider making
such changes.

> Software package that recognizes AGENTS.md would guide LLM to recognize 
> it. Yet without any guarantee. Testing something like that (multiple 
> LLMs) is not easy process. You would need to have GPU running and know 
> how to load models.

How do you know Sean didn't do just that?

> If you have tested remote LLMs with some software agent most probably 
> you tested those proprietaries. Yet those are just few among millions. I 
> have tested many LLMs running on my computer and they will not behave 
> same. You can get files deleted. Database can get deleted. There is no 
> guarantee with LLMs neither. I have gone through all that, so now I 
> learned how to behave so that those dangerous activities never get 
> invoked.

Once again: the fact this file will not always be effective is well
known.  Can you please stop beating this dead horse?

> >> Or we are just creating AGENTS.md for commercial software using LLM?!
> > 
> > Again, we created it to help users making innocent mistakes, using
> > whatever LLM-based agents, both proprietary and free.  The Emacs
> > policy is not to accept code generated by LLM-based agents regardless
> > of whether they are free or not.  As was repeatedly explained here.
> 
> AGENTS.md is not to help users. It is by specification to guide software 
> agents, not users.

And by doing that, it helps users to avoid LLM-generated code in their
contributions when they don't mean that.  So yes, it's eventually
helping users.

> Users are guided by README or CONTRIBUTE.

If they read those files and remember all of what they say (which is
not easy: CONTRIBUTE is more than 500 lines of terse technical text),
then yes.  The file AGENTS.md is for those who don't read or don't
remember or inadvertently invoke LLM-based agents in a way that
generates code when they didn't mean to.

Again, this has been explained many times.

> I was all the time thinking that Emacs policy is not to accept LLM 
> generated code.

It is.

> Now you said it is the policy not to accept LLM-based agent generated 
> code.

No, I didn't say that.  It's just that the model itself doesn't
produce anything, it should be used by some agent to do so.  IOW, it's
a matter of accurate terminology, that's all.

> Please, clarify it, is it fine then to generate code by LLM if we do not 
> use LLM-based agents?

No, code generated with LLM assistance will not be accepted, as long
as the current policy is in effect.  As was already explained many
times here.