That blog post is based on a study that also said:
>We conclude that while context files are useful for specifying non-standard coding practices, any attempts to improve performance should be rigorously evaluated before deployment
The kernel project has some particular conventions and workflows, so I think in this case it may make sense.
This is a load of shit. AGENTS.md is just a rules file for a repo like anything else. Properly-written rules files give good direction into how to interact with a project. If you don’t think it’s being read, ask the agent, and figure out why.
Far too many of these LLM articles are:
I did the thing
The thing didn’t work for me
Pretend it’s a hyper-scientific experiment and write an article about it
Declare X is shit because my version of X is shit, or because most people don’t know how to make X.
You don’t do the thing and pretend the first draft is going to work perfectly. You iterate, ask the agent directly why it didn’t work or what kind of improvements could be made, and improve the process.
Most of the time when I’m running through a repetitive process with a skill, I ask, “hey, what pitfalls did we learn from this process and how can we improve this skill?” Boom, iteration.
Also, Pivot-to-AI is just some guy’s anti-AI blog site. He does the same shit on YouTube.
It’s literally researchers doing research which the blog links to. Most of the AI booster stuff is completely unsubstantiated, and it turns out when you measure things they are often full of shit.
Bad research is still bad. The main paper is 9 pages with wide margins, outside of references. This was a few college students publishing their CS research assignment. The “Limitations” section is rather telling:
Another interesting avenue opened by this work is how to improve the automatic generation of useful context files.
Okay, so they already admitted that they started with useless context files, which sabotages their own test.
You know how you fix useless context files? You iterate until they are useful context files!
This whole research paper is XKCD 385, except with AGENTS.md files.
I haven’t read the paper in dept, but these are deeply unserious reasons for disparaging it.
Bad research is still bad. The main paper is 9 pages with wide margins, outside of references. This was a few college students publishing their CS research assignment
The length of this paper entirely within the norm for scientific publications. In fact, depending on the type of paper, they can be much shorter. For example, Nature expects submissions at around 2500 words or around 4300 words, depending on the format, while this paper is around 4800 words long (excluding references and SI).
The authors also obviously not “a few college students publishing their CS research assignment”: Gloaguen, Mündler, Raychev, Müller, and Vechev.
The “Limitations” section is rather telling:
Another interesting avenue opened by this work is how to improve the automatic generation of useful context files.
Okay, so they already admitted that they started with useless context files, which sabotages their own test.
You know how you fix useless context files? You iterate until they are useful context files!
The entire point of the study was to see what kind of context files were useful and which were useless, which according to the authors was not something that had been done before:
A widespread practice in software development is to tailor coding agents to repositories using context files, such as AGENTS.md. Although this practice is strongly encouraged by agent developers, there is currently no rigorous investigation into whether such context files are actually effective for real-world tasks. In this work, we study this question and evaluate coding agents’ task completion performance in two complementary settings
They found that,
while instructions in the context files are well followed by coding agents, repository overviews, although popular and recommended by model providers, are not helpful.
So when you say that they “admitted that they started with useless context files”, you are referring to the results of the study. Complaining that they didn’t base their study on the conclusions of that very study is just absurd
Please add the cost : this is not for everyone, not even close. Investor are paying for 99% of the cost right now with the hope of having everyone dependent on the tools to then multiply the prices by 100 once the “market penetration” feels satisfactory.
Also purely stealing people’s jobs from those same people’s work even when the work is explicit stating they refuse to be part of this farce.
Yeah I had one trying to start an initiative to add them and when I asked how they know if they’re doing anything he was completely at a loss for words. It is just cargo cult mentality.
https://pivot-to-ai.com/2026/08/27/your-agents-md-file-doesnt-actually-do-anything/
That blog post is based on a study that also said:
>We conclude that while context files are useful for specifying non-standard coding practices, any attempts to improve performance should be rigorously evaluated before deployment
The kernel project has some particular conventions and workflows, so I think in this case it may make sense.
They found it mostly just increases token spend.
This is a load of shit.
AGENTS.mdis just a rules file for a repo like anything else. Properly-written rules files give good direction into how to interact with a project. If you don’t think it’s being read, ask the agent, and figure out why.Far too many of these LLM articles are:
You don’t do the thing and pretend the first draft is going to work perfectly. You iterate, ask the agent directly why it didn’t work or what kind of improvements could be made, and improve the process.
Most of the time when I’m running through a repetitive process with a skill, I ask, “hey, what pitfalls did we learn from this process and how can we improve this skill?” Boom, iteration.
Also, Pivot-to-AI is just some guy’s anti-AI blog site. He does the same shit on YouTube.
It’s literally researchers doing research which the blog links to. Most of the AI booster stuff is completely unsubstantiated, and it turns out when you measure things they are often full of shit.
Bad research is still bad. The main paper is 9 pages with wide margins, outside of references. This was a few college students publishing their CS research assignment. The “Limitations” section is rather telling:
Okay, so they already admitted that they started with useless context files, which sabotages their own test.
You know how you fix useless context files? You iterate until they are useful context files!
This whole research paper is XKCD 385, except with
AGENTS.mdfiles.I haven’t read the paper in dept, but these are deeply unserious reasons for disparaging it.
The length of this paper entirely within the norm for scientific publications. In fact, depending on the type of paper, they can be much shorter. For example, Nature expects submissions at around 2500 words or around 4300 words, depending on the format, while this paper is around 4800 words long (excluding references and SI).
The authors also obviously not “a few college students publishing their CS research assignment”: Gloaguen, Mündler, Raychev, Müller, and Vechev.
The entire point of the study was to see what kind of context files were useful and which were useless, which according to the authors was not something that had been done before:
They found that,
So when you say that they “admitted that they started with useless context files”, you are referring to the results of the study. Complaining that they didn’t base their study on the conclusions of that very study is just absurd
Ai is shit because of the people behind it, and because of the crypto bro mentality of its users.
Please add the cost : this is not for everyone, not even close. Investor are paying for 99% of the cost right now with the hope of having everyone dependent on the tools to then multiply the prices by 100 once the “market penetration” feels satisfactory.
Also purely stealing people’s jobs from those same people’s work even when the work is explicit stating they refuse to be part of this farce.
Environmental costs, and the already recorded lower education scores as well.
ai was something that always seemed like a requirement for the future, but has now become the thing most likely to stop there being a future.
Coworkers have been adding these to our projects and I have noticed no difference in the usefullness of LLMs, so I suppose it’s nice to be validated.
Yeah I had one trying to start an initiative to add them and when I asked how they know if they’re doing anything he was completely at a loss for words. It is just cargo cult mentality.