The research mostly points to LLM-generated context lowering performance. Human-generated context improves performance, but any kind of AGENTS.md file increases token use, for what they say is "fake thinking." More research is needed.
I suspect AGENTS.md files will prove to be a short-lived relic of an era when we had to treat coding agents like junior devs, who often need explicit instructions and guardrails about testing, architecture, repo structure, etc. But when agents have the equivalent (or better) judgement ability as a senior engineer, they can make their own calls about these aspects, and trying to "program" their behaviour via an AGENTS.md file becomes as unhelpful as one engineer trying to micro-manage another's approach to solving a problem.
I never use these files and give the current guardrails of a specific task to each short run for agents. Have task specific “agents.md” works better for me.
I have a legacy codebase of around 300k lines spread across 1.5k files, and have had amazing success with the agents.md file.
It just prevents hallucinations and coerces the AI to use existing files and APIs instead of inventing them. It also has gold-standard tests and APIs as examples.
Before the agents file, it was just chaos of hallucinations and having to correct it all the time with the same things.
I liked they did this work + its sister paper, but disliked how it was positioned basically opposite of the truth.
The good: It shows on one kind of benchmark, some flavors of agentically-generated docs don't help on that task. So naively generating these, for one kind of task, doesn't work. Thank you, useful to know!
The bad: Some people assume this means in general these don't work, or automation can't generate useful ones.
The truth: Instruction files help measurably, and just a bit of engineering enables you to guarantee high scores for the typical cases. As soon as you have an objective function, you can flip it into an eval, and set an AI coder to editing these files until they work.
Ex: We recently released https://github.com/graphistry/graphistry-skills for more easily using graphistry via AI coding, and by having our authoring AI loop a bit with our evals, we jumped the scores from 30-50% success rate to 90%+. As we encounter more scenarios (and mine them from our chats etc), it's pretty straight forward to flip them into evals and ask Claude/Codex to loop until those work well too.
We do these kind of eval-driven AI coding loops all the time , and IMO how to engineer these should be the message, not that they don't work on average. Deeper example near the middle/end of the talk here: https://media.ccc.de/v/39c3-breaking-bots-cheating-at-blue-t...
I find this is very heavily situation dependent. I tried getting even Opus to implement a xero integration using the python sdk, without telling it much about the sdk besides the example code/instructions on github and it made an absolute un-runnable mess out of it.
The sdk and examples appear to be very atypical for a lot of python I've done (and atypical for my code base). Giving it a context/how-to file I put together made it significantly better (however I still gave up and did it more manually).
The trace analysis is the most interesting part of this paper.
Agents dutifully follow AGENTS.md instructions — they run more tests, grep more files, do more checks — but that thoroughness costs 19% more inference without meaningfully improving outcomes. That's not an argument against context files, it's evidence we're writing them wrong.
We're authoring them like READMEs for humans ("here's the architecture, here's how we structure things") when agents actually need narrow, non-inferable directives — the custom build flag, the weird test harness, the constraint they'd waste 30 tool calls discovering.
The HN commenter who noted the real value is "forcing you to articulate things previously just in your head" is actually pointing at the same conclusion from the other direction: the thinking matters, but the artifact needs a different shape than what we've converged on.
12 comments
[ 2.9 ms ] story [ 43.6 ms ] threadAGENTS.md are extremely helpful if done well.
It just prevents hallucinations and coerces the AI to use existing files and APIs instead of inventing them. It also has gold-standard tests and APIs as examples.
Before the agents file, it was just chaos of hallucinations and having to correct it all the time with the same things.
The good: It shows on one kind of benchmark, some flavors of agentically-generated docs don't help on that task. So naively generating these, for one kind of task, doesn't work. Thank you, useful to know!
The bad: Some people assume this means in general these don't work, or automation can't generate useful ones.
The truth: Instruction files help measurably, and just a bit of engineering enables you to guarantee high scores for the typical cases. As soon as you have an objective function, you can flip it into an eval, and set an AI coder to editing these files until they work.
Ex: We recently released https://github.com/graphistry/graphistry-skills for more easily using graphistry via AI coding, and by having our authoring AI loop a bit with our evals, we jumped the scores from 30-50% success rate to 90%+. As we encounter more scenarios (and mine them from our chats etc), it's pretty straight forward to flip them into evals and ask Claude/Codex to loop until those work well too.
We do these kind of eval-driven AI coding loops all the time , and IMO how to engineer these should be the message, not that they don't work on average. Deeper example near the middle/end of the talk here: https://media.ccc.de/v/39c3-breaking-bots-cheating-at-blue-t...
The sdk and examples appear to be very atypical for a lot of python I've done (and atypical for my code base). Giving it a context/how-to file I put together made it significantly better (however I still gave up and did it more manually).
Agents dutifully follow AGENTS.md instructions — they run more tests, grep more files, do more checks — but that thoroughness costs 19% more inference without meaningfully improving outcomes. That's not an argument against context files, it's evidence we're writing them wrong.
We're authoring them like READMEs for humans ("here's the architecture, here's how we structure things") when agents actually need narrow, non-inferable directives — the custom build flag, the weird test harness, the constraint they'd waste 30 tool calls discovering.
The HN commenter who noted the real value is "forcing you to articulate things previously just in your head" is actually pointing at the same conclusion from the other direction: the thinking matters, but the artifact needs a different shape than what we've converged on.