I read the study. I think it does the opposite of what the authors suggest - it's actually vouching for good AGENTS.md files.
> Surprisingly, we observe that developer-provided files only marginally improve performance compared to omitting them entirely (an increase of 4% on average), while LLM- generated context files have a small negative effect on agent performance (a decrease of 3% on average).
This "surprisingly", and the framing seems misplaced.
For the developer-made ones: 4% improvement is massive! 4% improvement from a simple markdown file means it's a must-have.
> while LLM- generated context files have a small negative effect on agent performance (a decrease of 3% on average)
This should really be "while the prompts used to generate AGENTS files in our dataset..". It's a proxy for prompts, who knows if the ones generated through a better prompt show improvement.
The biggest usecase for AGENTS.md files is domain knowledge that the model is not aware of and cannot instantly infer from the project. That is gained slowly over time from seeing the agents struggle due to this deficiency. Exactly the kind of thing very common in closed-source, yet incredibly rare in public Github projects that have an AGENTS.md file - the huge majority of which are recent small vibecoded projects centered around LLMs. If 4% gains are seen on the latter kind of project, which will have a very mixed quality of AGENTS files in the first place, then for bigger projects with high-quality .md's they're invaluable when working with agents.
Hey thanks for your review, a paper author here.
Regarding the 4% improvement for human written AGENTS.md: this would be huge indeed if it were a _consistent_ improvement. However, for example on Sonnet 4.5, performance _drops_ by over 2%. Qwen3 benefits most and GPT-5.2 improves by 1-2%.
The LLM-generated prompts follow the coding agent recommendations. We also show an ablation over different prompt types, and none have consistently better performance.
But ultimately I agree with your post. In fact we do recommend writing good AGENTS.md, manually and targetedly. This is emphasized for example at the end of our abstract and conclusion.
Without measuring quality of output, this seems irrelevant to me.
My use of CLAUDE.md is to get Claude to avoid making stupid mistakes that will require subsequent refactoring or cleanup passes.
Performance is not a consideration.
If anything, beyond CLAUDE.md I add agent harnesses that often increase the time and tokens used many times over, because my time is more expensive than the agents.
In this context, "performance" means "does it do what we want it to do" not "does it do it quickly". Quality of output is what they're measuring, speed is not a consideration.
CLAUDE.md isn't a silver bullet either, I've had it lose context a couple of questions deep. I do like GSD[1] though, it's been a great addition to the stack. I also use multiple, different LLMs as a judge for PRs, which captures a load of issues too.
> The biggest usecase for AGENTS.md files is domain knowledge that the model is not aware of and cannot instantly infer from the project. That is gained slowly over time from seeing the agents struggle due to this deficiency.
This. I have Claude write about the codebase because I get tired of it grepping files constantly. I rather it just know “these files are for x, these files have y methods” and I even have it breakdown larger files so it fits the entire context window several times over.
Funnily enough this makes it easier for humans to parse.
My pet peeve with AI is that it tends to work better in codebase where humans do well and for the same reason.
Large orchestration package without any tests that relies on a bunch of microservices to work? Claude Code will be as confused as our SDEs.
This in turns lead to broader effort to refactor our antiquated packages in the name of "making it compatible with AI" which actually means compatible with humans.
In Theory There Is No Difference Between Theory and Practice, While In Practice There Is.
In large projects, having a specific AGENTS.md makes the difference between the agent spending half of its context window searching for the right commands, navigating the repo, understanding what is what, etc., and being extremely useful. The larger the repository, the more things it needs to be aware of and the more important the AGENTS.md is. At least that's what I have observed in practice.
This reads a lot like bargaining stage. If agentic AI makes me a 10 times more productive developer, surely a 4% improvement is barely worth the token cost.
> If agentic AI makes me a 10 times more productive
I'm not sure what you are suggesting exactly, but wanted to highlight this humongous "if".
It's not only about the token cost! It's also my TIME cost! Much-much more expensive than tokens, it turns out ;)
If something makes you 10x as effective and then you improve that thing by 4%...
10x is that quantity or quality?
Also, "perceived" or "real"?
Honestly, the more research papers I read, the more I am suspicious. This "surprisingly" and other hyperbole is just to make reviewers think the authors actually did something interesting/exciting. But the more "surprises" there are in a paper, the more I am suspicious of it. Often such hyperbole ought to be at best ignored, at worst the exact opposite needs to be examined.
It seems like the best students/people eventually end up doing CS research in their spare time while working as engineers. This is not the case for many other disciplines, where you need e.g. a lab to do research. But in CS, you can just do it from your basement, all you need is a laptop.
Well, you still need time (and permission from your employer)! Research is usually a more than full time job on its own.
4% is yuuuge. In hard projects, 1% is the difference between getting it right with an elegant design or going completely off the rails.
The study measures the wrong thing. Task completion ("does the PR pass tests?") is a narrow proxy for what AGENTS.md actually helps with in production.
I run a system with multiple AI agents sharing a codebase daily. The AGENTS.md file doesn't exist to help the agent figure out how to fix a bug. It exists to encode tribal knowledge that would take a human weeks to accumulate: which directory owns what, how the deploy pipeline works, what patterns the team settled on after painful debates. Without it, the agent "succeeds" at the task but produces code that looks like it was written by someone who joined the team yesterday. It passes tests but violates every convention.
The finding that context files "encourage broader exploration" is actually the point. I want the agent to read the testing conventions before writing tests. I want it to check the migration patterns before creating a new table. That costs more tokens, yes. But reverting a merged PR that used the wrong ORM pattern costs more than 20% extra inference.
What are you putting in the file? When I’ve looked at them they just looked like a second readme file without the promotional material in a typical GitHub readme.
The useful stuff is different from a README. A README tells humans how to use the project. An AGENTS.md tells the AI how to work on it.
Mine typically includes:
- Build/test commands that aren't obvious from package.json (e.g. "run migrations before tests") - Architecture decisions that would take the agent 10 minutes to reverse-engineer ("auth goes through middleware X, not controller Y") - Known gotchas ("don't touch the legacy billing module, it's being replaced next sprint") - Deploy process specifics ("push to main auto-deploys staging, prod needs a manual tag") - Coding conventions that aren't in the linter ("we use Result types for errors, never throw")
The ones that look like READMEs are indeed useless. The good ones read more like the notes you'd give a new senior engineer on their first day. Stuff that's obvious to the team but invisible to an outsider.
That's basically all it is. It's a readme file that is guaranteed to be read. So the agent doesn't spend 10 minutes trying to re-configure the toolchain because the first command it guessed didn't work.
This is why I only add information to AGENTS.md when the agent has failed at a task. Then, once I've added the information, I revert the desired changes, re-run the task, and see if the output has improved. That way, I can have more confidence that AGENTS.md has actually improved coding agent success, at least with the given model and agent harness.
I do not do this for all repos, but I do it for the repos where I know that other developers will attempt very similar tasks, and I want them to be successful.
You can also save time/tokens if you see that every request starts looking for the same information. You can front-load it.
Also take the randomness out of it. Sometimes the agent executing tests one way, sometimes the other way.
I've found https://github.com/casey/just to be very very useful. Allows to bind common commands simple smaller commands that can be easily referenced. Good for humans too.
Don't forget to update it regularly then
Agree. I also found out that rule discovery approach like this perform better. It is like teaching a student, they probably have already performed well on some task, if we feed in another extra rule that they already well verse at, it can hinder their creativity.
That's a sensible approach, but it still won't give you 100% confidence. These tools produce different output even when given the same context and prompt. You can't really be certain that the output difference is due to isolating any single variable.
So true! I've also setup automated evaluations using the GitHub Copilot SDK so that I can re-run the same prompt and measure results. I only use that when I want even more confidence, and typically when I want to more precisely compare models. I do find that the results have been fairly similar across runs for the same model/prompt/settings, even though we cannot set seed for most models/agents.
same with people, no matter what info you give a person you cant be sure they will follow it the same every time
I've found that even documenting non-obvious dependencies between tasks can significantly improve agent performance and reduce debugging time
When an agent just plows ahead with a wrong interpretation or understanding of something, I like to ask them why they didn't stop to ask for clarification. Just a few days ago, while refactoring minor stuff, I had an agent replace all sqlite-related code in that codebase with MariaDB-based code. Asked why that happened, the answer was that there was a confusion about MariaDB vs. sqlite because the code in question is dealing with, among other things, MariaDB Docker containers. So the word MariaDB pops up a few times in code and comments.
I then asked if there is anything I could do to prevent misinterpretations from producing wild results like this. So I got the advice to put an instruction in AGENTS.md that would urge agents to ask for clarification before proceeding. But I didn't add it. Out of the 25 lines of my AGENTS.md, many are already variations of that. The first three:
- Do not try to fill gaps in your knowledge with overzealous assumptions.
- When in doubt: Slow down, double-check context, and only touch what was explicitly asked for.
- If a task seems to require extra changes, pause and ask before proceeding.
If these are not enough to prevent stuff like that, I don't know what could.
Are agents actually capable of answering why they did things? An LLM can review the previous context, add your question about why it did something, and then use next token prediction to generate an answer. But is that answer actually why the agent did what it did?
It depends. If you have an LLM that uses reasoning the explanation for why decisions are made can often be found in the reasoning token output. So if the agent later has access to that context it could see why a decision was made.
Reasoning, in majority of cases, is pruned at each conversation turn.
The cursor-mirror skill and cursor_mirror.py script lets you search through and inschpekt all of your chat histories, all of the thinking bubbles and prompts, all of the context assembly, all of the tool and mcp calls and parameters, and analyze what it did, even after cursor has summarized and pruned and "forgotten" it -- it's all still there in the chat log and sqlite databases.
cursor-mirror skill and reverse engineered cursor schemas:
https://github.com/SimHacker/moollm/tree/main/skills/cursor-...
cursor_mirror.py:
https://github.com/SimHacker/moollm/blob/main/skills/cursor-...
The German Toilet of AI
"The structure of the toilet reflects how a culture examines itself." — Slavoj Zizek
German toilets have a shelf. You can inspect what you've produced before flushing. French toilets rush everything away immediately. American toilets sit ambivalently between.
cursor-mirror is the German toilet of AI.
Most AI systems are French toilets — thoughts disappear instantly, no inspection possible. cursor-mirror provides hermeneutic self-examination: the ability to interpret and understand your own outputs.
What context was assembled?
What reasoning happened in thinking blocks?
What tools were called and why?
What files were read, written, modified?
This matters for:
Debugging — Why did it do that?
Learning — What patterns work?
Trust — Is this skill behaving as declared?
Optimization — What's eating my tokens?
See: Skill Ecosystem for how cursor-mirror enables skill curation.
https://news.ycombinator.com/item?id=23452607
According to Slavoj Žižek, Germans love Hermeneutic stool diagnostics:
https://www.youtube.com/watch?v=rzXPyCY7jbs
>Žižek on toilets. Slavoj Žižek during an architecture congress in Pamplona, Spain.
>The German toilets, the old kind -- now they are disappearing, but you still find them. It's the opposite. The hole is in front, so that when you produce excrement, they are displayed in the back, they don't disappear in water. This is the German ritual, you know? Use it every morning. Sniff, inspect your shits for traces of illness. It's high Hermeneutic. I think the original meaning of Hermeneutic may be this.
https://en.wikipedia.org/wiki/Hermeneutics
>Hermeneutics (/ˌhɜːrməˈnjuːtɪks/)[1] is the theory and methodology of interpretation, especially the interpretation of biblical texts, wisdom literature, and philosophical texts. Hermeneutics is more than interpretive principles or methods we resort to when immediate comprehension fails. Rather, hermeneutics is the art of understanding and of making oneself understood.
----
Here's an example cursor-mirror analysis of an experiment with 23 runs with four agents playing several turns of Fluxx per run (1 run = 1 completion call), 1045+ events, 731 tool calls, 24 files created, 32 images generated, 24 custom Fluxx cards created:
Cursor Mirror Analysis: Amsterdam Fluxx Championship -- Deep comprehensive scan of the entire FAFO tournament development:
amsterdam-flux CURSOR-MIRROR-ANALYSIS.md:
https://github.com/SimHacker/moollm/blob/main/skills/experim...
amsterdam-flux simulation runs:
https://github.com/SimHacker/moollm/tree/main/skills/experim...
Just an update re German toilets: No toilet set up in the last 30 years (I know of) uses a shelf anymore. This reduces water usage by about 50% per flush.
But then what do you have to talk about all day??!
Well, the entire field of explainable AI has mostly thrown in the towel..
of course not, but it can often give a plausible answer, and it's possible that answer will actually happen to be correct - not because it did any - or is capable of any - introspection, but because it's token outputs in response to the question might semi-coincidentally be a token input that changes the future outputs in the same way.
Isn't that question a category error? The "why" the agent did that is that it was the token that best matched the probability distribution of the context and the most recent output (modulo a bit of randomness). The response to that question will, again, be the tokens that best match the probability distribution of the context (now including the "why?" question and the previous failed attempt).
if the agent can review its reasoning traces, which i think is often true in this era of 1M token context, then it may be able to provide a meaningful answer to the question.
Wait, no, that's the category error I'm talking about. Any answer other than "that was the most likely next token given the context" is untrue. It is not describing what actually happened.
I think this statement is on the same level as "a human cannot explain why they gave the answer they gave because they cannot actually introspect the chemical reactions in their brain." That is true, but a human often has an internal train of thought that preceded their ultimate answer, and it is interesting to know what that train of thought was.
In the same way, it is often quite instructive to know what the reasoning trace was that preceded an LLM's answer, without having to worry about what, mechanically, the LLM "understood" about the tokens, if this is even a meaningful question.
But it's not a reasoning trace. Models could produce one if they were designed to (an actual stack of the calls and the states of the tensors with each call, probably with a helpful lookup table for the tokens) but they specifically haven't been made to do that.
When you put an LLM in reasoning mode, it will approximately have a conversation with itself. This mimics an inner monologue.
That conversation is held in text, not in any internal representation. That text is called the reasoning trace. You can then analyse that trace.
Unless things have changed drastically in the last 4 months (the last time I looked at it) those traces are not stored but reconstructed when asked. Which is still the same problem.
They aren't necessarily "stored" but they are part of the response content. They are referred to as reasoning or thinking blocks. The big 3 model makers all have this in their APIs, typically in an encrypted form.
Reconstruction of reasoning from scratch can happen in some legacy APIs like the OpenAI chat completions API, which doesn't support passing reasoning blocks around. They specifically recommend folks to use their newer esponses API to improve both accuracy and latency (reusing existing reasoning).
For a typical coding agent, there are intermediate tool call outputs and LLM commentary produced while it works on a task and passed to the LLM as context for follow up requests. (Hence the term agent: it is an LLM call in a loop.) You can easily see this with e.g. Claude Code, as it keeps track of how much space is left in the context and requires "context compaction" after the context gradually fills up over the course of a session.
In this regard, the reasoning trace of an agent is trivially accessible to clients, unlike the reasoning trace of an individual LLM API call; it's a higher level of abstraction. Indeed, I implemented an agent just the other day which took advantage of this. The OP that you originally replied to was discussing an agentic coding process, not an individual LLM API call.
Well, right, I see those reasoning stages in reasoning models with Ollama and if you ask it what its reasoning was after the fact what it says is different than what it said at the time.
There can be higher- and lower-level descriptions of the same phenomenon. when the kettle boils, it’s because the water molecules were heated by the electric element, but it’s also because I wanted a cup of tea.
the llm has no wants
> Any answer other than "that was the most likely next token given the context" is untrue.
"Because the matrix math resulted in the set of tokens that produced the output". "Because the machine code driving the hosting devices produced the output you saw". "Because the combination of silicon traces and charges on the chips at that exact moment resulted in the output". "Because my neurons fired in a particular order/combination".
I don't see how your statement is any more useful. If an LLM has access to reasoning traces it can realistically waddle down the CoT and figure out where it took a wrong turn.
Just like a human does with memories in context - does't mean that's the full story - your decision making is very subconscious and nonverbal - you might not be aware of it, but any reasoning you give to explain why you did something is bound to be an incomplete story, created by your brain to explain what happened based on what it knows - but there's hidden state it doesn't have access to. And yet we ask that question constantly.
well, do you want something useful or something true?
the word why is used to get something true.
If you want to be pedantic about it you could phrase it as follows.
When the LLM was in reasoning mode, in the reasoning context it often expressed statement X. Given that, and the relevance of statement X to the taken action. It seems likely that the presence of statement X in the context contributed to this action. Besides, the presence of statement X in the reasoning likely means that given the previous context embeddings of X are close to the context.
Hence we think that the action was taken due to statement X.
And that output could have come from an LLM introspecting it's own reasoning.
I don't think that phrasing things so pedanticaly is worth the extra precision though. Especially not for the statement that inspecting the reasoning logs of sn LLM can help give insight on why an LLM acted a certain way.
Just this morning I have run across an even narrower case of how AGENTS.md (in this case with GPT-5.3 Codex) can be completely ignored even if filled with explicit instructions.
I have a line there that says Codex should never use Node APIs where Bun APIs exist for the same thing. Routinely, Claude Code and now Codex would ignore this.
I just replaced that rule with a TypeScript-compiler-powered AST based deterministic rule. Now the agent can attempt to commit code with banned Node API usage and the pre-commit script will fail, so it is forced to get it right.
I've found myself migrating more and more of my AGENTS.md instructions to compiler-based checks like these - where possible. I feel as though this shouldn't be needed if the models were good, but it seems to be and I guess the deterministic nature of these checks is better than relying on the LLM's questionable respect of the rules.
I wonder if some of these could be embedded in the write tool calls?
Not that much different from humans.
We have pre-commit hooks to prevent people doing the wrong thing. We have all sorts of guardrails to help people.
And the “modern” approach when someone does something wrong is not to blame the person, but to ask “how did the system allow this mistake? What guardrails are missing?”
I really hate that the anthropomorphizing of these systems has successfully taken hold in people's brains. Asking it why it did something is completely useless because you aren't interrogating a person with a memory or a rationale, you’re querying a statistical model that is spitting out a justification for a past state it no longer occupies.
Even the "thinking" blocks in newer models are an illusion. There is no functional difference between the text in a thought block and the final answer. To the model, they are just more tokens in a linear sequence. It isn't "thinking" before it speaks, the "thought" is the speech.
Treating those thoughts as internal reflection of some kind is a category error. There is no "privileged" layer of reasoning happening in the silicon that then gets translated into the thought block. It’s a specialized output where the model is forced to show its work because that process of feeding its own generated strings back into its context window statistically increases the probability of a correct result. The chatbot providers just package this in a neat little window to make the model's "thinking" part of the gimmick.
This is like trying to fix hallucination by telling LLM not to hallucinate.
> So I got the advice to put an instruction in AGENTS.md that would urge agents to ask for clarification before proceeding.
You may want to ask the next LLM versions the same question after they feed this paper through training.
It seems like LLMs in general still have a very hard time with the concepts of "doubt" and "uncertainty". In the early days this was very visible in the form of hallucinations, but it feels like they fixed that mostly by having better internal fact-checking. The underlying problem of treating assumptions as truth is still there, just hidden better.
Doubt and uncertainty is left for us humans.
LLMs are basically improv theater. If the agent starts out with a wildly wrong assumption it will try to stick to it and adapt it rather than starting over. It can only do "yes and", never "actually nevermind, let me try something else".
I once had an agent come up with what seemed like a pointlessly convoluted solution as it tried to fit its initial approach (likely sourced from framework documentation overemphasizing the importance of doing it "the <framework> way" when possible) to a problem for which it to me didn't really seem like a good fit. It kept reassuring me that this was the way to go and my concerns were invalid.
When I described the solution and the original problem to another agent running the same model, it would instantly dismiss it and point out the same concerns I had raised - and it would insist on those being deal breakers the same way the other agent had dimissed them as invalid.
In the past I've often found LLMs to be extremely opinionated while also flipping their positions on a dime once met with any doubt or resistance. It feels like I'm now seeing the opposite: the LLM just running with whatever it picked up first from the initial prompt and then being extremely stubborn and insisting on rationalizing its choice no matter how much time it wastes trying to make it work. It's sometimes better to start a conversation over than to try and steer it in the right direction at that point.
so many times have ended up here :
"You're absolutely correct. I should have checked my skills before doing that. I'll make sure I do it in the future."
My personal experience is that it’s worthwhile to put instructions, user-manual style, into the context. These are things like:
- How to build.
- How to run tests.
- How to work around the incredible crappiness of the codex-rs sandbox.
I also like to put in basic style-guide things like “the minimum Python version is 3.12.” Sadly I seem to also need “if you find yourself writing TypeVar, think again” because (unscientifically) it seems that putting the actual keyword that the agent should try not to use makes it more likely to remember the instructions.
I also try to avoid negative instructions. No scientific proof, just a feeling the same as you, "do not delete the tmp file" can lead too often to deleting the tmp file.
It’s like instructing a toddler.
i definitely have gone so far as to treat my llm readable docs in this way and have found it very effective
I recall that early LLMs had the problem of not understanding the word "not", which became especially evident and problematic when tasked with summarizing text because the summary would then sometimes directly contradict the original text.
It seems that that problem hasn't really been "fixed", it's just been paved over. But I guess that's the ugly truth most people tend to forget/deny about LLMs: you can't "fix" them because there's not a line of code you can point to that causes a "bug", you can only retrain them and hope the problem goes away. In LLMs, every bug is a "heisenbug" (or should that be "murphybug", as in Murphy's Law?).
For TypeVar I’d reach for a lint warning instead.
I also have felt like these kinds of efforts at instructions and agent files have been worthwhile, but I am increasingly of the opinion that such feelings represent self-delusion from seeing and expecting certain things aided by a tool that always agrees with my, or its, take on utility. The agent.md file looks like it’d work, it looks how you’d expect, but then it fails over and over. And the process of tweaking is pleasant chatting with supportive supposed insights and solutions, which means hours of fiddling with meta-documentation without clear rewards because of partial adherence.
The papers conclusions align with my personal experiments at managing a small knowledge base with LLM rules. The application of rules was inconsistent, the execution of them fickle, and fundamental changes in processing would happen from week-to-week as the model usage was tweaked. But, rule tweaking always felt good. The LLM said it would work better, and the LLM said it had read and understood the instructions and the LLM said it would apply them… I felt like I understoood how best to deliver data to the LLMs, only to see recurrent failures.
LLMs lie. They have no idea, no data, and no insights into specific areas, but they’ll make pleasant reality-adjacent fiction. Since chatting is seductive, and our time sense is impacted by talking, I think the normal time versus productivity sense is further pulled out of ehack. Devs are notoriously bad at estimating where they’re using time, long feedback loops filled with phone time and slow ass conversation don’t help.
Yesterday while i was adding some nitpicks to a CLAUDE.md/AGENTS.md file, I thought « this file could be renamed CONTRIBUTING.md and be done with it ».
Maybe I’m wrong but sure feels like we might soon drop all of this extra cruft for more rationale practices
Exactly my thoughts... the model should just auto ingest README and CONTRIBUTING when started.
You could have claude --init create this hook and then it gets into the context at start and resume
Or create it in some other way
{
"hookSpecificOutput": {
"hookEventName": "SessionStart",
"additionalContext": "<contents of your file here>"
}
}And that makes total sense. Honestly working since a few days with Opus 4.6, it really feels like a competent coworker, but need some explicit conventions to follow … exactly when onboarding a new IC! So i think there is a bright light to be seen: this will force having proper and explicit contribution rules and conventions, both for humans and robots
Quite a surprising result: “across multiple coding agents and LLMs, we find that context files tend to reduce task success rates compared to providing no repository context, while also increasing inference cost by over 20%.”
Well, task == Resolving real GitHub Issues
Languages == Python only
Libraries (um looks like other LLM generated libraries -- I mean definitely not pure human: like Ragas, FastMCP, etc)
So seems like a highly skewed sample and who knows what can / can't be generalized. Does make for a compelling research paper though!
Hey, paper author here. We did try to get an even sample - we include both SWE-bench repos (which are large, popular and mostly human-written) and a sample of smaller, more recent repositories with existing AGENTS.md (these tend to contain LLM written code of course). Our findings generalize across both these samples. What is arguably missing are small repositories of completely human-written code, but this is quite difficult to obtain nowadays.
Why stick to python-only repositories though?
To reduce the number of variables to account for. To be able to finish the paper this year, and not the next century. To work with a familiar language and environments. To use a language heavily represented in the training data.
I mean, it's not that hard to understand why.
> Libraries (um looks like other LLM generated libraries -- I mean definitely not pure human: like Ragas, FastMCP, etc)
How does this invalidate the result? Aren't AGENTS.md files put exactly into those repos that are partly generated using LLMs?
I think that is a rather fitting approach to the problem domain. A task being a real GitHub issue is a solid definition by any measure, and I see no problem picking language A over B or C.
If you feel strongly about the topic, you are free to write your own article.
LLMs are generally bad at writing non-noisy prompts and instructions. It's better to have it write instructions post hoc. For instance, I paste this prompt into the end of most conversations:
If there’s a nugget of knowledge learned at any point in this conversation (not limited to the most recent exchange), please tersely update AGENTS.md so future agents can access it. If nothing durable was learned, no changes are needed. Do not add memories just to add memories.
Update AGENTS.md **only** if you learned a durable, generalizable lesson about how to work in this repo (e.g., a principle, process, debugging heuristic, or coding convention). Do **not** add bug- or component-specific notes (for example, “set .foo color in bar.css”) unless they reflect a broader rule.
If the lesson cannot be stated without referencing a specific selector or file, skip the memory and make no changes. Keep it to **one short bullet** under an appropriate existing section, or add a new short section only if absolutely necessary.
Another common mistake is to have very long AGENTS.md files. The file should not be long. If it's longer than 200 lines, you're certainly doing it wrong.
we've been running AGENTS.md in production on helios (https://github.com/BintzGavin/helios) for a while now.
each role owns specific files. no overlap means zero merge conflicts across 1800+ autonomous PRs. planning happens in `.sys/plans/{role}/` as written contracts before execution starts. time is the mutex.
AGENTS.md defines the vision. agents read the gap between vision and reality, then pull toward it. no manager, no orchestration.
we wrote about it here: https://agnt.one/blog/black-hole-architecture
agents ship features autonomously. 90% of PRs are zero human in the loop. the one pain point is refactors. cross-cutting changes don't map cleanly to single-role ownership
AGENTS.md works when it encodes constraints that eliminate coordination. if it's just a roadmap, it won't help much.
Most of these AI-guiding "techniques" seem more like reading into tea leaves to me than anything actually useful.
Even with the latest and greatest (because I know people will reflexively immediately jump down my throat if I don't specify that, yes, I've used Opus 4.6 and Gemini 3 Pro etc. etc. etc. etc., I have access to all of the models by way of work and use them regularly), my experience has been that it's basically a crapshoot that it'll listen to a single one of these files, especially in the long run with large chats. The amount of times I still have to tell these things to not generate React in my Vue codebase that has literally not a single line of JSX anywhere and instructions in every single possible file I can put it in to NOT GENERATE FUCKING REACT CODE makes me want to blow my brains out every time it happens. In fact it happened to me today with the supposed super intelligence known as Opus 4.6 that has 18 trillion TB of context or whatever in a fresh chat when I asked for a quick snippet I needed to experiment with.
I'm not even paying for this crap (work is) and I still feel scammed approximately half the time, and can't help but think all of these suggestions are just ways to inflate token usage and to move you into the usage limit territory faster.
Claude/Opus 4.6 Can you add a console.log in food XYZ?
No problem, x agents, hundreds/closed to one million token usage to add a line of code.
Gemini 3 : can you review the commit A (console.log one ) you have made the most significant change in your 200kloc code base, this key change will allow you to get great insight into your software.
Codex : I have reviewed your change, you are missing tests and integration tests.
But I fully agree, overall I feel there are a lot of tea leaves readers online and LinkedIn.
I'd be interested to see results with Opus 4.6 or 4.5
Also, I bet the quality of these docs vary widely across both human and AI generated ones. Good Agents.md files should have progressive disclosure so only the items required by the task are pulled in (e.g. for DB schema related topics, see such and such a file).
Then there's the choice of pulling things into Agents.md vs skills which the article doesn't explore.
I do feel for the authors, since the article already feels old. The models and tooling around them are changing very quickly.
Agree that progressive disclosure is fantastic, but
> (e.g. for DB schema related topics, see such and such a file).
Rather than doing this, put another AGENTS.md file in a DB-related subfolder. It will be automatically pulled into context when the agent reads any files in the file. This is supported out of the box by any agent worth its salt, including OpenCode and CC.
IMO static instructions referring an LLM to other files are an anti-pattern, at least with current models. This is a flaw of the skills spec, which refers to creating a "references" folder and such. I think initial skills demos from Anthropic also showed this. This doesn't work.
This is probably the best comment in the thread. I've totally forgotten about nested AGENTS.md files, gonna try implementation today.
Progressive disclosure is good for reducing context usage but it also reduces the benefit of token caching. It might be a toss-up, given this research result.
Those are different axes - quality vs money.
Progressive disclosure is invaluable because it reduces context rot. Every single token in context influences future ones and degrades quality.
I'm also not sure how it reduces the benefit of token caching. They're still going to be cached, just later on.
It is still baffling to me why we need AGENTS.md
Any well-maintained project should already have a CONTRIBUTING.md that has good information for both humans and agents.
Sometimes I actually start my sessions like this "please read the contributing.md file to understand how to build/test this project before making any code changes"
Just symlink CONTRIBUTING as AGENTS
This only works on systems that support symlinks. It also pollutes the root folder with more files.
I understand the sentiment, but it is really strange that the people that are pushing for agents.md haven't seen https://contributing.md/
Is it even mentioned at GitHub docs https://docs.github.com/en/communities/setting-up-your-proje...
If the harnesses had a simple system prompt "read repository level markdown and honor the house style"?
Think of the agent app store people's children man, it would be a sad Christmas.
What is the purpose of an AGENTS.md file when there are so many different models? Which model or version of the model is the file written for? So much depends on assumptions here. It only makes sense when you know exactly which model you are writing for. No wonder the impact is 'all over the place'.
Their definition of context excludes prescriptive specs/requirements files. They are only talking about a file that summarizes what exists in the codebase, which is information that's otherwise discoverable by the agent through CLI (ripgrep, etc), and it's been trained to do that as efficiently as possible.
Also important to note that human-written context did help according to them, if only a little bit.
Effectively what they're saying is that inputting an LLM generated summary of the codebase didn't help the agent. Which isn't that surprising.
I find it surprising. The piece of code I'm working on is about 10k LoC to define the basic structures and functionality and I found Claude Code would systematically spend significant time and tokens exploring it to add even basic functionality. Part of the issue is this deals with a problem domain LLMs don't seem to be very well trained on, so they have to take it all in, they don't seem to know what to look for in advance.
I went through a couple of iterations of the CLAUDE.md file, first describing the problem domain and library intent (that helped target search better as it had keywords to go by; note a domain-trained human would know these in advance from the three words that comprise the library folder name) and finally adding a concise per-function doc of all the most frequently used bits. I find I can launch CC on a simple task now, without it spending minutes reading the codebase before getting started.
That's also my experience.
The article is interesting but I think it deviates from a common developer experience as many don't work on Python libraries, which likely heavily follow patterns that the model itself already contains.
Hey, a paper author here :) I agree, if you know well about LLMs it shouldn't be too surprising that autogenerated context files are not helping - yet this is the default recommendation by major AI companies which we wanted to scrutinize.
> Their definition of context excludes prescriptive specs/requirements files.
Can you explain a bit what you mean here? If the context file specifies a desired behavior, we do check whether the LLM follows it, and this seems generally to work (Section 4.3).
I only put things when the LLM gets something wrong and I need to correct it. Like “no, we create db migrations using this tool” kind of corrections. So far it made them behave correctly in those situations.
I'd take any paper like this with a grain of salt. I imagine what holds true for models in time period X could drastically be different just given a little more time.
Doesn't mean it's not worth studying this kind of stuff, but this conclusion is already so "old" that it's hard to say it's valid anymore with the latest batch of models.
This is life of an LLM researcher. We literally ran the last experiments only a month ago on what were the latest models back then...
In my experience AGENTS.md files only save a bit of time, they don't meaningfully improve success. Agents are smart enough to figure stuff out on their own, but you can save a few tool calls and a bit of context by telling them how to build your project or what directories do what rather than letting it stumble its way there.
Check the logs, no one really requests AGENTS.md from the server.
This paper shoulda just done a study on elixirs usage_rules?
The only thing I use CLAUDE.md for is explaining the purpose and general high level design principles of the project so I don't have to waste my time reiterating this every time I clear the context. Things like this is a file manager, the deliverable must always be a zipapp, Wayland will never be supported.
I added these to that file because otherwise I will have to tell claude these things myself, repeatedly. But the science says... Respectfully, blow it out your ass.
I pretty much add to my prompt bunch of stuff, with AGENST.md or any file I can just add one line "hey read up that file".
If I understand the paper correctly, the researchers found that AGENTS.md context files caused the LLMs to burn through more tokens as they parsed and followed the instructions, but they did not find a large change in the success rate (defined by "the PR passes the existing unit tests in the repo").
What wasn't measured, probably because it's almost impossible to quantify, was the quality of the code produced. Did the context files help the LLMs produce code that matched the style of the rest of the project? Did the code produced end up reasonably maintainable in the long run, or was it slop that increased long-term tech debt? These are important questions, but as they are extremely difficult to assign numbers to and measure in an automated way, the paper didn't attempt to answer them.
Research has shown that most earlier "techniques" to get better LLM response no longer work and are actively harmful with modern models. I'm so grateful that there's actual studies and papers about this and that they keep coming out. Software developers are super cargo culty and will do whatever the next guy does (and that includes doing whatever is suggested in research papers)
Software developers don't have to be cargo-culty... if they're working on systems that are well-documented or are open-source (or at least source-available) so that you can actually dig in to find out how the system works.
But with LLMs, the internals are not well-documented, most are not open-source (and even if the model and weights are open-source, it's impossible for a human to read a grid of numbers and understand exactly how it will change its output for a given input), and there's also an element of randomness inherent to how the LLM behaves.
Given that fact, it's not surprising to find that developers trying to use LLMs end up adding certain inputs out of what amounts to superstition ("it seems to work better when I tell it to think before coding, so let's add that instruction and hopefully it'll help avoid bad code" but there's very little way to be sure that it did anything). It honestly reminds me of gambling fallacies, e.g. tabletop RPG players who have their "lucky" die that they bring out for important rolls. There's insufficient input to be sure that this line, which you add to all your prompts by putting it in AGENTS.md, is doing anything — but it makes you feel better to have it in there.
(None of which is intended as a criticism, BTW: that's just what you have to do when using an opaque, partly-random tool).
Many of the practices in this field are mostly based on feelings and wishful thinking, rather than any demonstrable benefit. Part of the problem is that the tools are practically nondeterministic, and their results can't be compared reliably.
The other part is fueled by brand recognition and promotion, since everyone wants to make their own contribution with the least amount of effort, and coming up with silly Markdown formats is an easy way to do that.
EDIT: It's amusing how sensitive the blue-pilled crowd is when confronted with reality. :)