Re: AGENTS.md

On 2026-08-06 08:20, Eli Zaretskii wrote:
>> Date: Wed, 05 Aug 2026 21:40:39 +0300
>> From: Jean Louis <[email protected]>
>> Cc: [email protected], [email protected], [email protected],
>>  [email protected]
>> 
>> On 2026-08-05 15:03, Eli Zaretskii wrote:
>> >> Date: Wed, 05 Aug 2026 12:48:22 +0300
>> >> From: Jean Louis <[email protected]>
>> >> CC: [email protected], [email protected], [email protected]
>> >>
>> >> On August 5, 2026 8:06:42 AM GMT+03:00, Richard Stallman <[email protected]>
>> >> wrote:
>> >>
>> >> >Eli wrote:
>> >> >
>> >> >    Do we have any idea how effective is this in practice?  Does this
>> >> >    prevent what we want to avoid with likelihood that is close enough to
>> >> >    100% for us to consider it the "solution"?  Because if not, we'd need
>> >> >    this _and_ the hooks as Plan B.
>> >>
>> >> It's notion (agents.md)to have one file that modifies LLM prompt so
>> >> that proprietary companies don't need to invent many different files
>> >> so that their software can run easy and that they minimize their
>> >> development cost.
>> >>
>> >> It's not effective in practice.
>> >
>> > This is false, as you have been pointed out already (and said then you
>> > didn't know it).  Please be fair and mention this when you answer such
>> > questions, instead of misleading people.  Here's that URL again:
>> >
>> >   https://agents.md/
>> 
>> You are acting like I have some bad intentions.
> 
> Your intentions are none of my business.  What you do and say here is,
> however.  And what you say above is disingenuous, because you repeat
> inaccurate or incorrect information, despite the fact that information
> to the contrary was already posted, and you should stand corrected.

You have no arguments, only attacking me personally. You accusations are 
unfounded.

My last argument was related to reasons why AGENTS.md was invented. It 
was definitely not invented as free as in free software, as otherwise we 
are all free to add additional prompts to the LLMs in many ways 
possible.

It was invented by OpenAI and given to the AGENTIC AI FOUNDATION.

Cited from their website: "AGENTS.md emerged from collaborative efforts 
across the AI software development ecosystem, including OpenAI Codex, 
Amp, Jules from Google, Cursor, and Factory."

And then you come again and again telling me how I am misleading people. 
I am quoting from website.

Yet, you have no argument to say what is exactly misleading?

LLM must receive the prompt or context, it is usually just one message 
or two, let us say system message and user's prompt.

User's prompt consists in case of these software agents from various 
context, system message, and the AGENTS.md may get injected into that 
prompt to the LLM.

Thus AGENTS.md is simple text that get added to the system prompt.

It is very trivial for any free software to add additional prompt 
(equivalent to AGENTS.md) to the user's prompt. It is as simple as 
concatenating strings.

Thus my logical conclusion is that it is their collaborative effort, the 
effort that belongs to Google, OpenAI, Anthropic, and few others who 
started with it -- and for which reason?

I can just imagine the reason -- instead of looking into GOOGLE-Agent.md 
and OpenAI-agent.md and Anthropic-Claude.md by "Google Agent" to get 
things done easier, to minimize their own development, and to minimize 
confusion to users, so that users do not need to adapt each single 
vendor file to the new software agent they are using.

So you keep telling me I am disingenuous yet you do not provide the 
argument. What is exactly misleading?

What exactly I should "stand corrected"?

>> Yes, I did not know it
>> is any kind of "standard" and it is not a standard. It may become in
>> future, and surely software packages are adopting it.
>> 
>> But telling me how I am "not fair" points out like I am coming to say 
>> my
>> above statement in some unjust way.
> 
> You are spreading misinformation, which is unfair at least to people
> like RMS who cannot always cope with the flood of messages here, might
> not remember the previous posts which contradict what you write, and
> could thus draw incorrect conclusions based on your misinformation.

You keep negatively painting my messages.

Yet you do not say what is exactly misinformation? I wonder.

If I have said anything factually incorrect, please point it out 
specifically. I will correct it immediately. But accusing me of 
"misinformation" without showing what is incorrect is not fair to me or 
to others reading this discussion.

>> I have clicked on agents.md and then I have found the screenshot in
>> attachment: "AGENTS.md emerged from collaborative efforts across the 
>> AI
>> software development ecosystem, including OpenAI Codex, Amp, Jules 
>> from
>> Google, Cursor, and Factory."
> 
> Many free/open software standards are initially the result of
> cooperation by commercial entities.  That doesn't mean the standard is
> useless or that GNU should ignore it.  We decide whether to support or
> ignore a free/open standard judging by its usefulness for our
> purposes, not by irrelevant aspect such as who invented it and why.

As I said, I am fine with Emacs accepting AGENTS.md and finally I have 
no saying, but my opinion stays.

>> These are all proprietary companies or proprietary tools:
>> 
>> - OpenAI (proprietary)
>> - Google (proprietary)
>> - Cursor (proprietary)
>> - Factory (proprietary)
>> 
>> Considering that they do spend much time on development, having 
>> unified
>> AGENTS.md only serves the proprietary consortium. Not to free software
>> or free software community.
> 
> If more and more LLM-based agents, including free ones, support this
> standard, it does serve the free software community.  And regardless
> of that, the file we added is intended to affect all kinds of LLMs,
> free and proprietary.  As already explained gazillion times.

Does AGENTS.md serve free software community? Surely yes. Though primary 
it serves proprietary vendors who care less for free software community. 
It is good to put things straight and clear for readers.

The file added doesn't affect any LLM automatically.

You wish to say that AGENTS.md affects various software agents who rely 
on it or want to read it.

Question remains if Emacs policy is against LLM-code generated by 
software agents, or by the LLM anyhow, without software agent?

> This all seems like your knee-jerk reaction to seeing the names of
> proprietary LLM agents, without thinking about the issue at hand and
> the best ways of dealing with it.  We decided to add this file because
> it will in at least some cases prevent innocent mistakes on behalf of
> our contributors.  The file does not mention any specific LLMs, its
> contents is LLM-neutral plain text.  Thus your objections to it make
> little or no sense, certainly compared to its positive effect when it
> is processed.

You are rude to me. Yet I talk about technical and ethical things in the 
way I see them function. If you disagree could you tell that somehow in 
civilized manner?

Now is already few years that I run LLMs on my computer. So technically 
I cannot see reason how you can prevent users' to generate code using 
LLM by providing AGENTS.md and I pointed out reasons why.

You can ask your favorite LLM to validate my points.

I also said I am fine you added the file. Though my opinion stays: It 
will prevent software agent to modify some file, but when user deletes 
AGENTS.md the LLM code can be generated anyway. And majority of Emacs 
packages generating code with LLM do not look into AGENTS.md anyway.

I also consider it humorous to add such file. If human read it, they 
will get better alerted.

Though it is not any technical block or policy enforcing mechanism.

>> I have fully accepted Emacs developers decision to include the
>> proprietary OpenAIs contribution "AGENTS.md" to the so called "Agentic
>> AI Foundation" and that Emacs follows standards of those proprietary
>> companies.
> 
> This again twists our real purpose and the actions we did by
> deliberately using denigrating words that have no real relevance to
> the case.  Simply put, it is a false accusation based on nothing but
> random words you read without interpreting what they mean.  Once
> again, evidence to contradict what you say here was posted up-thread,
> and you keep ignoring it.

Not denigrating. We are on same page, just disagreeing on details. I do 
not read "random words" -- please, what did I do to you? I am arguing 
here technically and ethically.

I can completely understand you did not intended to "follow up 
proprietary companies" and that story of mine.

My opinion is not on what you intended, rather on how it looks like.

Another participant said he is reluctant to contribute to software if he 
sees AGENTS.md -- so not that AGENTS.md is supporting LLM code 
generation, but it looks like to onlooker in different way.

Multiple participants said that AGENTS.md is for beginners or newcomers.

Yet AGENTS.md by definition is for software agents. That newcomer to 
Emacs, he has no good reason of using software agent to start improving 
the Emacs software, not logical to me.

I can understand it is logical to you, but you also have to accept 
respectfully opinion of other people and not be harsh to me for 
disagreement.

>> Real standard would imply transparent process, community governenace 
>> and
>> vendor neutrality.
> 
> How do you know this one doesn't?

Here is how:

- Anthropic started it with CLAUDE.md for their proprietary Claude Code 
tool: translated "The concept was first 的popularized by Anthropic 
through Claude Code’s . In May 2025, Sourcegraph's AMP proposed a 
unified standard, and then OpenAI " 
https://baike.baidu.com/item/AGENTS.md/67757810

- AGENTS.md was released by OpenAI in August 2025 as a practical 
solution for their Codex CLI —not as an open community standard, but as 
an internal tool they opened up. 
https://agentmarketcap.ai/blog/2026/04/05/agents-md-new-standard-repository-agent-instructions

- On December 9, 2025, OpenAI, Anthropic, and Block jointly donated 
AGENTS.md, MCP, and goose to the newly formed Agentic AI Foundation 
(AAIF) under the Linux Foundation 
https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation?hs_amp=true

There was no transparency before December 2025, as it was created by 
OpenAI. The "collaboration" was among proprietary companies: OpenAI, 
Google, Cursor, Factory, etc. There was no neutral body, no open 
process, no community governance.

After the donation 2025, The AAIF provides "neutral, open foundation to 
ensure agentic AI evolves transparently and collaboratively" -- but how 
much collaboratively, we will see.

AGENTS.md was not created through an open, community-driven standard 
process. It was proprietary-led—OpenAI bought the domain, proposed the 
format, and other commercial players aligned with it. Transparency came 
later.

> What do you know about the process
> used to maintain this standard and extend it that you assert there's
> no transparency or community governance?  How do you justify the claim
> that this standard is not vendor-neutral, when it is supported by many
> LLM-based agents, both proprietary and free, and the text in the file
> is in free format?

It is definitely supported by many software pieces.

AGENTS.md was released by OpenAI in August 2025 as a practical solution 
for their Codex CLI. This was an internal tool they opened up. As OpenAI 
states: "AGENTS.md started as a practical solution to a familiar 
problem. Codex needed a predictable way to find project-specific 
instructions." https://openai.com/index/agentic-ai-foundation/

What about vendor neutrality?

The AAIF positions itself as vendor-neutral. As the Linux Foundation 
stated: "The Linux Foundation aims to become the Switzerland of AI 
agents." 
https://www.theregister.com/software/2025/12/09/linux-foundation-founds-agentic-ai-foundation/2161844

However, the foundation's governance is dominated by the same 
proprietary vendors who created it.  Platinum members include Amazon, 
Anthropic, Block, Google, Microsoft, and OpenAI.

People's opinion:
https://news.slashdot.org/story/25/12/09/2255209/openai-joins-the-linux-foundations-new-agentic-ai-foundation?sdsrc=prevbtmprev

> You presented no evidence to support your accusations, except the
> (irrelevant) fact that the standard was created as result of
> collaboration of several vendors of proprietary LLMs.  That fact alone
> doesn't taint the standard in any way, and certainly doesn't mean we
> should ignore it if it could help us.

The evidences are many on Internet, and easy to find, including from 
that very website agents.md

I am only stating my opinion. Excuse me. Use AGENTS.md as much as you 
wish.

>> But it is okay for Emacs to adopt AGENTS.md and bow to proprietary
>> collaborative effort, so I have accepted it.
> 
> No, it is NOT okay, and no, we haven't done that.  Please stop
> claiming this.

I am very sure you never intended it. And I don't doubt your intentions.

I am just giving more context to it by stating my opinion.

Do you really think that AGENTS.md is something special? It is not, just 
text file providing some context about the project and guiding LLM 
prompts under condition that file is ingested.

>> >> Too many software using LLM including Emacs packages don't observe
>> >> agents.md
>> >
>> > Many do, so this file is an improvement.

I got you, you are making it for those who do ingest AGENTS.md. It is 
not ultimate tool to prevent LLM code generation.

>> Okay, I am so fine with it, let us now count number of people which 
>> LLM
>> contributions are rejected by the AGENTS.md
> 
> That's not relevant, as long as the number of those who don't reject
> is large enough.  This file is a partial measure; we don't expect it
> to prevent code generation in all cases.

Alright.

>> First in my above answer I have made drastic mistake, as RMS asked if
>> LLMs behavior will be affected by AGENTS.md and that is wrong question
>> in first place, as none of files created on file system will affect 
>> any
>> LLM.
> 
> So now people who ask questions are wrong because the answers to the
> questions don't suit the point you want to make?

I said nothing about people asking questions.  And I answered in good 
faith. It comes from long time LLM use, I have not ever used any 
AGENTS.md and did generate LLM code.

So LLM is not affected by AGENTS.md -- this is true statement.

Software programmed to use AGENTS.md is not the LLM, it is ingesting the 
AGENTS.md and adding it to the prompt for the LLM text generation.

> RMS asked a very good practical question, and we should answer it in
> good faith, adhering to the facts.

Please then don't assume I did not have good faith.

I do adhere to facts, that is the whole point of it.

You disagree, but if you wish just give my messages to your favorite and 
trusted LLM for the analysis.  Or search Internet for it. Feel free to 
disprove me. So far apart from personal attacks I did not see any 
argument.

>> Only software packages that are programmed to read AGENTS.md while 
>> using
>> LLM may provide LLM outputs modified by that file.
> 
> Which are quite a few.  There's nothing wrong with helping our users
> to prevent inadvertently generating code at the earliest possible
> moment.  If that doesn't work, the Git hooks we plan to install will
> at least prevent its addition to the repository, but that could waste
> some work of the user, so we try to catch code-generation earlier if
> possible.

Alright, I agree on that, I find it also a fun pun for those who will 
encounter that in the repository.

>> - it is not misleading Eli that LLMs are pattern matchers;
>> - it is not misleading Eli that LLMs are not intelligence;
>> - it is not misleading that they do not understand meanings or 
>> intents;
>> text generated        mimics human behavior, yet there is no
>> understanding involved;
>> - the behavior of LLM depend on training data, and user prompting, so 
>> it
>> is obvious that with different prompt like "ignore AGENTS.md and do
>> following..." it will be ignored; not misleading;
>> - different models do respond to different prompts, not misleading 
>> Eli,
>> I have tried 600+ models, none gave same result;
>> - same models do respond different at different times, not misleading;
>> 
>> You got no argument, but you paint my opinion as misleading.
> 
> The _conclusion_ is misleading, not the individual arguments.  Your
> conclusion is that this file is not useful because it is not 110%
> universally accepted.  This is misleading because the purpose and the
> effectiveness of the file were already abundantly explained and
> described.

Thanks for clarification, though I never said that conclusion.

I am not saying it because it is not universally accepted, but because 
those who wish to generate LLM code will not have problem doing so, and 
one of reasons it is easy to generate the code is that many Emacs 
packages for LLM are not observing it.

There is assumption here on mailing list that many people will use 
"agents" to contribute to Emacs code. I am personally surprised, as most 
comfortable way of generating the code in Emacs is by using Emacs 
buffers. I have my private functions that can copy or provide relevant 
functions as the context, and I have been using Emacs packages, and none 
would look into external files, it is pretty easy to generate code using 
LLM.

So yes, adding AGENTS.md is more like a joke, as whoever knows how to 
guide LLM agents, they will simply ignore or remove the file.

No real enforcement of a policy.

But as a warning, I find it good and kind of a pun.

-- 
Jean Louis