It is most certainly the kind of subtle bug the LLMs love to slip in and burn you in production

Not sure what y’all are thinking with these unwritten conspiracy theories that begin and end with “it’s intentional”

Get more people to turn on telemetry.

I like to give my graces to people and the companies who are typically not trillions of dollars. Have an incredible amount of resources that many countries would like to have, with fewer of the obligations. I'm going to chalk this up to its intended

Someone on the Claude Code team is probably wondering the same...

many of these can also be chalked up to the fact that all of their products are extremely vibed

I cannot set a password for login, on logging in it says we've sent a login code but it's a link instead...

Vibe-coding at its best.

Nice find, though I'd rather read the prompt that was used to write this article. It is about ten times longer than it needs to and is quite painful to read.

Yep, especially with long contexts (and moreso if the last thing you were working on in the same context involved telemetry too). AI sneaks in weird conditions like this and then does the entire "You're absolutely right" thing if you're paying enough attention to catch it.

No, people seem to be adopting simplistic narratives and then slotting the facts wherever they most conveniently fit into their narrative.

One thing I do know is that an Anthropic employee is definitely NOT going to blame this on the model they are using.

Except that, and this I find slightly amusing, is a situation where they definitely don't want to publicly blame the ai model when there are mistakes.

I used to actively use Msty for local models because it just worked and had a lot of nice advanced features. A few months ago they released their beta version of a Claw-like UI and mentioned using a version of it to develop Msty itself. Well, that was around the time I stopped using Msty because every update started breaking things and two updates in a row included bugs which wiped all my configs, chats, and history.

ps: apparently thats not how you spell emblematic - but I'm actually glad to leave a little humanity around given the topic :D

Or it’s as simple as “hey, they don’t want telemetry? Cool, we’ll turn off all the extra phoning home” and their feature flag code uses the same paths as telemetry because it’s the same endpoint, etc.

Oops, easy mistake!

Good thing they don't apply the same standards to security sandboxes.

Turns out AGI boils down to two markdown files:

claude-md-or-agents-md

claude-md-and-agents-md

This is what's written in their release notes for 2.1.277:

Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead

My dumb brain tells me none of this is rolled out progressively (as of that version). You either have it or not.

It's been like this at least months. I turned telemetry off a few months ago and it basically prevented all stages rollouts of new features from working.

Anthropic commonly gates behind feature flags that require telemetry until they're "promoted" and defaulted on.

A little bit annoying you can't manually control the flags without telemetry but I think the title is a bit click bait.

I use cursor and claude, its kinda annoying having to have the same skills in both so I made a ~/.agents/skills folder then ln both cursor and claude skills to point to that folder.

which works except that claude uses .skills/synced which is uses to sync changes to skills from claude servers into the skills folder.

every other agent I have used just directly syncs into .skills so it ends up duplicating skills

can i ask claude to change these settings for me, and will it enable telemetry as implied checkpoint when i ask it to enable AGENTS.md?

It's not open source. It's more "source available", since it's published, but you legally can't do anything with it. Other than maybe build it yourself, for yourself, I guess.

I think you are not missing anything. I have not checked the code itself originally, I am sorry for that. But you are right that there seem to be no actual code for the harness itself. Just the code for (some) mods, plugins and scripts.

I just read the first line in the README file which says:

Claude Code is an agentic coding tool ...

and I immediately assumed that this is what this repository hosts.

It claims that it includes plugins but that does not mean it does not include anything else. It also never explicitly claims, as far as I can tell, that it does not hold the source code of Claude Code, the harness, itself.

It is all extremely misleading, in my opinion. Which might be on purpose, unfortunately.

That's not the source code. Claude Code is proprietary.

Uhm I turned off everything there is to turn off in my Pro plan, and Claude just read my agents.md with no issue? So... What telemetry am I missing? Or did they JUST update? (I updated my docker image 40 minutes ago to get Opus 5.5, am on version 2.1.280, so not the mentioned 2.1.277, so it's fixed?)

Much respect for the prompt response, humility, and frank disclosure.

I don't understand, isn't coding solved already?

One line change? Pffft. That means you're still looking at the code, you're behind the times.

I think you're giving these files too much credit. There are no specs, they are freeform text. In this sense they are the same. The expectation of what could be in it by each vendor means nothing unless it's enforced.

It's not a format though is it? it's literally just an extension to the system prompt in plain markdown.

There is no rhyme or reason to the structure of this file, just like with most things in AI. It's best effort human language.

No, they do not.

Nope! Claude is even able to understand AsciiDoc just fine (which is a blessing). But it will always insist on the Markdown extension.

There is no "specification", it's literally just a text file that gets loaded into context. You don't even need to write markdown. It's not that hard man.

Yes, but at the same time, it’s also a good, simple use case to test a new plugin system. I can totally understand that.

https://youtu.be/-AwCEM5K-Ec

Unc you're so cooked it's not even funny younger than me but you're out here aura farming in the negatives fr fr. NPC-ahh behavior, 0 aura, -1000 rizz. Couldn't be me

Sorry couldn't resist

Not sure what you didn't understood in their comment though, maybe the raph loop (and jev which was just released)? Isn't the rest completely self explanatory?

It's almost all related to LLM's, which is fairly new.

Basically he's describing an LLM assistant setup that does coding on a vps or Mac mini ( some contradiction with Openrouter though) where you can communicate with the assistant through telegram

I've never heard of the 'move fast, break things' mentality ever giving a damn about number of users. If so, no changes would ever be made. The great thing about having millions of users is QA can be eliminated entirely as you'll start hearing about issues from the users directly.

What's your point? Nobody can hold themselves to a higher standard because they didn't in the past?

Bash does this well, reading only the first of half a dozen config file locations it searches. And if the user needs it to read from an additional location, just add a source line to that location.

Because when you have tons of users ainor fuckuo is a big fuckup and also it's really common to have both claude.md and agents.md and use @ syntax (which lets you reference markdown files when using Claude code, but not other harnesses) so you Claude md looks like

```md @AGENTS.md

[Claude specific stuff]

```

And then what happens if someone now puts @syntax in their agents.md triggering a loop etc. It's all vibe coded - including code from days with dumber models - there's gonna be all sorts of dragons under the hood

What, exactly, is going to go horribly wrong here ...?

Operating anything at scale teaches you that the universe of “things that can go horribly wrong” has a much better imagination than you. That goes quadruple with complex and squishy reasoning models.

I’ve seen a couple of really good examples of what can go wrong in this subthread, the most straightforward of which is that abruptly changing a fundamental behavior can break existing use cases.

I would like to subscribe to your newsletter.

Seriously, though, this is a great idea. Good communicators know to tailor their discourse to their audience. LLMs are just as sensitive to small variations in prompts, and just as inconsistent between them, as are humans.

It's not a bad idea, although at that point it probably shouldn't be a .md file anymore. You'd want to use YAML or something.

I did this at one point with Jinja templates.

I wrote an agent launcher sort of bash script. Pass in the command to start the agent, the script checks if there’s a Jinja file in a special directory matching that name, and builds it to AGENTS.md. Then it launches the agent.

I was trying to use it as a sort of janky RAG. I had a bunch of snippets (one for DB architecture, one for how load balancing works, etc), and my Jinja files were mostly a list of snippets to pull in. Voila, a bunch of agents that share little pieces of info but have a single source of truth.

I never got a ton of value tbh, it was very good at just grepping the snippets.

This "tiny feature" consists of hundreds of lines of code spread across 57 source files, many of which are less than 30 lines long, often contain more comments than code, and have no real reason to exist. It's the very definition of overengineering and overcomplexity.

That said, I do find it somewhat funny that Anthropic themselves are falling victim to their own model's tokenmaxxing.

The Claude sandbox has a slightly different ENV compared to ChatGPT (LC_CTYPE set to C instead of UTF8 I believe) which upsets some scripts we have.

So in the CLAUDE.MD we tell it about that and then tell it continue reading AGENTS.MD. Not very hard to work around if CLAUDE.MD goes away, but just to support your point that the split did have its uses for some.

  > Having a kill switch for the change is a perfectly reasonable safety measure in case something goes horribly wrong.

Having a way for the user to revert the update is a perfectly reasonable safety measure in case something goes horribly wrong.

Even though Anthropic or any other software provider could technically run anything on our system, there exists a modicum of trust between supplier and user. Anthropic performing arbitrary changes that are tangential but not specific to the premise of delivering a harness to package code files and prompts in context to their LLM breaks that trust for many users.

With apologies for not just testing this myself (currently AFK), doesn't it still work to have a CLAUDE.md file containing just `@AGENTS.md`?

Do you really need a whole plugin system

Vibe-coding gotta vibe-code.

Go download the leaked source code from earlier this year, search for all occurrences of the string `CLAUDE.md`, and be horrified.

I'd love to hear more about this. Though I usually write bespoke software for specific clients, I'm about to release my first general-availability application. It's a rust application with Python GUI on the desktop (Linux, Mac, Windows) and a skinny Android application that wraps the rust on mobile. I'm going to be gating features starting probably next week. I'd love to hear tips and experience for implementing feature flags, for both the desktop and mobile space.

Thanks!

Welcome to SWE in 2026

need to, no, but if you put the Ai to task, expect over engineering

Sounds like you've read the deepseek harness paper.

Will you extend your plugin to read skills and rules from `.agents`? Or should we write our own plugin/mod for that?

If the direct link isn't working for anyone else (looks like it's some AWS pre-signed one maybe tied to IP): https://github.com/user-attachments/files/31802150/EXTERNAL....

they are noun phrases and not sentences at all? I’m pretty sure that any decent pre-LLM NLP-based grammar checker would correctly flag half the sentences in Claudish. Also, whatever variant of Claude wrote this thing can’t even capitalize around semicolons consistently with itself, let alone consistently with how English has been written for at least a century.

Probably just a case of a company hoping their scale can change the societal standard faster than they can be bothered to match the standard.

You'll talk like 2023 unsupervised TikTok generators and you'll be happy.

Your comment here is written in perfectly fine English :)

Maybe "function_hook_modules" for the JSON key?

Does this finally fix the limitation that there is no supportable way to replace the implementation of a native tool in Claude Code? codex-rs has an (undocumented but likely supportable if kind of awkward) mechanism, and I’m part way through a project that needs this. I was going to release it with a recommendation not to use Claude Code, but it’s at least plausible that this will change my mind. (I’m not sure yet — the $ mechanism might be too restrictive. It’s a start at least.)

Wow, you're Xe Iaso! I've reached out.

via our new extensibility system for CC, called Mods, which is launching soon-ish. A mod is a plugin with a new type of hook, which we call a function hook.

An extensibility system called mods, which is a plugin with a new type of hook that we call function hook?

I can't tell if this is real, or you are making fun of overengineered AI solutions.

Is this real?

mods seem like a grasp at all the pi and dsh users

I've developed several plugins for different harnesses, and I needed some upstream change for most of them.

The deciding factor for me whether or not I will work on the feature of the plugin is whether I (or rather, my agent) can look in upstream source and evaluate if it can be done with minimal upstream change, which I then contribute. And generally, even if no upstream change is needed, agents work so much better when they can read the code.

So why not just make Claude code open source? Considering also that source code was leaked once anyway.

Right... you can just do it yourself.

There are two or three relevant companies in this space in America and this is the one of them that kicked off the whole terminal agent harness thing in getting market adoption. It's perfectly fine for neither of these companies to follow industry standards while they're figuring shit out

After I was mildly disappointed with GPT-6 Sol and Luna not improving intelligence and only cutting the price, I'm running Opus 5.5 today.

After the last month or so in the Codex app, I was pleased with the Claude app.

It might be a case of the grass always being greener on the other side, but this is what stands out:

After 3-4 hours of usage, the weekly usage limit moved by only 1%.

Compared to Astra where I can watch the limit draining live, this is a great improvement.

I'd estimate it 3x cheaper, and that's with a 450k context limit instead of the 258k in Codex.

So far Opus 5.5 appears less prone to stopping for no apparent reason at checkpoints in the middle of a longer task.

It doesn't open an internal browser with a useless comparison page, where it then proceeds to add notes despite no one having asked for it.

It is a breath of fresh air: I get the response in the chat, while the Codex app recently loves randomly opening artifacts instead.

Opus 5.5 xhigh made great progress on the task, more so than Astra High, but that could be random chance.

Oh, and the 'Auto' mode actually works and does not force me to instead run 'Full access' like in the Codex app, lest it blocks even 'git push'.

Sure a fast response on HN would make people switch. Try better rates, infra, limits etc.

What we need is a low level but constant drumbeat against openai in general. In general the AI situation is overleveraged and underpoliced, with the occasional hints of AI gone wild. If openai were to just be left to die, we could let that financial mess unroll and bail out the leftovers, I don't like bailouts anymore than the next guy but with this administration its almost a guarantee if things go south because this adminstration can charge administrative fees of maybe $20-30 billion (which goes to trump), get Sam Altman to serve one or two years in a cushy resort type fed place for the hugging face hacking and put openai's processes on github as a premium feature, say $10000 a month to access (which again goes to trump).

I know I know, why are we giving money to trump? Its because he's going to take it anyways so can't we at least apply some window dressing?

Wow, thanks for sharing your experience. Very insightful.

Somehow this just hasn't happened to me. I use Astra on high all day for fairly intensive game dev tasks, sometimes cranked up depending on the task.

I have heard ultra thinking might delegate to worse agents for some of its sub-tasks, but I don't use that much anymore since Astra came out. Just high seems good enough to throw most laundry lists at.

It's night and day difference, for 3 days, until Anthropic nerf the model.

...if Anthropic can keep it up with this level of service...

fuckin laughable, literally invoked a laugh from me in real life.

I hope customers aren't so stupid that they think a chatty developer on twitter/hn/mastodon/screaming-in-the-wind/wherever (or any other public-facing-place) means shit about customer service, and that goes towards ANY company where the primary customer service is an LLM.

Anthropic is the only company where it took (!) 9 weeks (!) to convince to hand over a 4 dollar refund for book-keeping errors on their side that caused an inappropriately early account deactivation due to time zone issues on their end, while all the while telling me that they don't offer refunds. It took stacks of evidence and argument, and that was after spending two weeks in their system trying to convince every level that I was worth a human.

For me personally it'd require Dario to resort to armed mugging to see another buck out of my wallet. I'm not alone.

tl;dr : being able to convince the powers that be on highly active industry forums (hacker news, twitter, mastodon..?) to act right using the power of peer shaming doesn't good customer service make. That said -- I do appreciate the direct response/statement from mpoteat;

..I just don't appreciate the good actions of a decent individual being too broadly interpreted as the do-good customer-centric nature of Anthropic .. an element I do not believe exists there.

The same guy writing readme's for my vibeslopped toy projects is also the readme writer at Anthropic, what a coincidence ;)

Well now, this is how you do community outreach

Ouch.

I've had to send such messages, but internally at work, not on HN!

Have a great day, human.

Or add one additional telemetry metric and an extra prompt

The correct design was in the AGENTS.md but they didn't have telemetry on

I hope disabling /r if telemetry is disabled is also unintentional...

I was probably an agent that made the change, and the same agent that commented here on HN.

Clearly not Claude-speak, see the rest of the message. "Rollout artifact" just means "an artifact of how we rolled this change out" which maybe isn't proper English but the message was obviously not written by Claude. Upgrade your Claudish detection

Feels pretty normal to me, IDK. "Rollout" is definitely what was happening here, and this leftover problem can be described as an "artifact" most generally - I guess otherwise it'd be a... just "problem" or "mistake"? Plus, the Claudism here would probably involve "soak", and possibly "wall-time" lol

IMHO it's worth keeping in mind that Anthropic employees are some of the least likely to casually pass off artificial prose as authentic, given the company's ethos/brand/cover story (depending on how cynical you are). To them this is all getting pretty high stakes pretty damn quickly; based on my usage of full strength Opus 5.5 today, I can't even imagine what working with their full internal stack must feel like. If they were willing to let the machines speak for them, they'd all be melancholically lounging around home by now instead of coming in to work!

...I am refusing to consider the fact that they probably are still WFH because of Salesforce forcing their shared security contractor to strike. Call that a mental health ignorance on my part :)

How are you planning to turn this off remotely when telemetry off?

people are overeacting

valid

Thanks for admitting to the mistake, but my understanding was that coding was fully solved now?

how strange.

as i said in another comment. i don't want someone to toy with my software remotely. that seems wrong to me!

i do not like others to decide that they know what is best for me. and then force it on me without my consent.

i will decide if i like your changes. if i do like your fix, i will install it.

in my car, do not remotely turn off my air conditioning. don't turn off my AGENTS.md.

Not necessarily. They probably have an integrated service that handles some telemetry and also feature flags, like Braze. They toggled the whole thing off based on telemetry settings. It's a bug I've made before too.

That's just how big companies roll out software changes for software that auto-updates.

It's much preferable to be able to instantly fix it if the rollout of a new feature goes wrong than have everyone who installed the broken version bring stuck with problems until the company realizes the issue and rolls forwards with a fixed version.

https://martinfowler.com/articles/feature-toggles.html

telling people Anthropic remote-control their users computers isn't the smartest thing to do

is there reason I can't update claude to 2.1.281? I just run claude update

claude update Current version: 2.1.280 Checking for updates to latest version... Claude Code is up to date (2.1.280)

My only edit was to point out at the start that this design is indefensible, and I added the reason why. The rest matches the original meaning. I don't think my stance changed between before and after the edit. What Claude Code did is simply wrong.

it is indefensible.

Claude Code also doesn't read AGENTS.md by default if there's a CLAUDE.md it can read instead. This isn't limited to your repo, e.g. if you have a ~/CLAUDE.md then no AGENTS.md will be read.

To always read both, you have to switch the 'Project instructions' setting to the non-default `claude-md-and-agents-md`.

Just in case anyone is wondering why their AGENTS.md still isn't being read.

This is not the only case where Anthropic has done stuff silently without giving users any information about changes that would hurt them.

The number of people raw-dogging software that executes arbitrary instructions on their machine coming from a 3rd party server just absolutely baffles me.

The same people who've spent years of their career making sure that never happens.

Urm... no tests caught this? How?

Yeah, no thank you. This was why I polished my setup for VPN -> SSH -> [MOSH] -> TMUX -> claude/codex

This in some way sounds like VSCode's bug of always adding Copilot as a co-author of git commit regardless of user settings.

If people only glance over the code agents generate for them and don't bother to spend even half a minute thinking through what's actually happening, this is inevitable.

Certainly this kind of things happened before LLMs existed. But I'm not optimistic about the direction of how things are going.

Anyway, my claude.md contains only this:

@agents.md

Related:

Claude Code now reads AGENTS.md if there is no Claude.md

https://news.ycombinator.com/item?id=49760187

And I’m sure there’s no conceivable way an organization as well resourced as Anthropic can separate out feature flags from telemetry. It’s just too complicated! Maybe when we get AGI?

You're absolutely right! AGENTS.md shouldn't be gated behind telemetry on

Meanwhile, Codex devs just merge their changes and hit the release button. If a thingy breaks, they fix the thingy and release a hotfix. How irresponsible! It’s a miracle the software works at all!

Exactly! I can not believe this is in the top spot.

If they can't even get this simple thing right, I am worried about the future of anthropic's products.

Wait wait wait WAIT

So, these tools have a file they want to read in with some configuration.

That filename is hardcoded?

Fucking DOOM had a command line parameter to provide an arbitrary configuration file name!

That's completely irrespective of the fact that you need a feature flag set by remote infrastructure to change a setting of which completely local file to read.

It's weird, I feel like Claude would have tried to make this a configurable setting by default! Is that just not an option in JS land? Not a common pattern to have configuration in the first place? I don't know about that, all the JS based code editors have comprehensive configuration files.

What the hell is going on....

Couldn't you just make a CLAUDE.md that says, "Read AGENTS.md in this same directory"?

Embarrassing for HN to have a huge thread over a Feature Flag.

https://martinfowler.com/articles/feature-toggles.html