I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.
— Boris Cherny, head of Claude Code
Reliability is a direct reflection of the quality of the underlying infrastructural code. If even Anthropic, the company with the world's best agentic vibecoders, has horribly unreliable infrastructure, it really says something about the quality of the world's best agentically produced code.
Is there any indication these errors are related to Anthropic-written code as opposed to operational issues from the fastest-growing infra buildout ever?
Layer-wise, the app is pretty far removed from request routing to GPU pools.
This is almost certainly a software issue, though. Even if it's due to scaling, they still built a system that failed catastrophically rather than degrading gracefully.
Sure. But could it be k8s config? Could it be Nvidia Bright Cluster? Could it be load balancing?
I'm not saying Anthropic isn't to blame for a system that is literally approaching one-nine uptime; they certainly are. I am saying that jumping to the "it must be vibe coding's fault" is an emotional confirmation-bias belief, not an evidence-based belief.
I'd expect that they're also managing their k8s config and other infra using LLMs (it's actually quite good at this, at least for my simple homelab use-cases).
Right. If this were truly a pure scaling issue, I’d expect the interface would offer an archive.is-esque “Claude is at capacity; your prompt is #XXX/YYY in the queue; estimated time remaining: ZZZ seconds”
Instead, the whole system just shits the bed, catastrophically.
But such messages would suggest that Claude has engineered limits, which isn't what the market wants to hear. Completely falling over and being unavailable is just another Tuesday on the internet, will be forgotten by the weekend.
> being unavailable is just another Tuesday on the internet, will be forgotten by the weekend.
This is true when you have like one failure a year, but Anthropic is starting to look a lot like github lately when it comes to uptime.
After a certain point the reputation for unreliability starts sticking to you, especially when you position yourself as an indispensable tool for completing work people need done.
I'm not sure if that's really an Anthropic problem you're pointing to vs a problem that their infra layer handles (Amazon, Google, whatever hyperscaler). i.e, they might be scaling quickly but they are running on top of established infrastructure.
I would bet that they have inference setup for internal use on a separate system from the customer-facing production environment. The same way telemetry infrastructure needs to be run separate from normal production systems, so you aren't "blind" when you need it most.
> If even Anthropic, the company with the world's best agentic vibecoders...
But that's really not what they have. They have AI experts who are creating incredible LLMs.
Everything else is more than meh: Claude Code is really bad. Such a turd would never have gained any traction if it wasn't for the LLMs behind it.
I use LLMs to code daily (Claude Code still, mind you, for I didn't take the time to switch yet) and these modesl are both amazing and pathetic.
If you don't verify everything they output, they do the absolute craziest thing imaginable.
One example is I got an Anthropic model notice a "pattern" in range bound integer values. I had them range bound between, e.g., 0xCAFE0000 and 0xCAFEFFFF. And at some point a comparison/validation was needed and instead of doing an integer comparison the Anthropic model went ballistic: instead of doing an integer comparison it converted the numbers to a string, then started doing substring matching on "0xCAFE" and went even more "expert" by verifying at which position the match was happening. All that while explaining why it couldn't possibly fail.
Why did it do that? Very likely because, in a comment, it saw "0xCAFE..." as a string. And the thing saw a pattern.
Can you believe it? There's a pattern. So it must light up connections. We've got a pattern!
Now amount of kludge, hidden pre-processing, hidden post-processing is fixing the "quality" of the code produced by something that, instead of doing an integer comparison, converts things to string and then does substring searches and indexes computation.
There's no fixing that.
Yesterday: had to use three guard clauses before pushing data... Two of the three "logic gates" (as the model would explain they were, which is kinda right) he got right. The third one: same thing... It was planning to go ballistic, introduce countless lines of code, insane abstractions, to make a test that was solved with a one line timestamp comparison.
It's because it does things like that that the people who explain that they don't code anymore are delusional if they think this gives, as of today, quality code.
It's like that other dude who was happy to produce 37 K LOC per day and counting.
> ... it really says something about the quality of the world's best agentically produced code
Oh it is totally shit code. But if you monitor everything and vet everything they do, it's helpful.
I find these LLMs way more helpful at finding the source of bugs (not fixing them: finding them, which is 90% of the job anyway) and at acting like rubber-ducks then at writing code.
Claude Code sucks. Claude Code CLI sucks. Their only "solutions" to all problems is to create VMs, headless browsers, and resort to incredible hacks (the infamous "game loop" that modifies the characters output by the LLM is just shameful) etc. to try to hide the misery. It's miserable kludges everywhere.
And the only reason these miserable kludges are not entirely falling apart is because they rest on the shoulders of actual giants: projects like Linux, QEMU, etc. that were not vibe-coded.
It's sad to have useful tools (the models) and to make such poor use of them.
I'm pretty sure that, in the end, it's just like open-source powering the entire world by now: we'll have open-source projects like Pi and then newer ones that are going to come out and fix the mess we have now. And they're not going to be 100% vibe-coded by people whose jobs is "to write loops".
He is a salesman at this point and is not talking to you. He is talking to the investors who want to vibe code loops to waste tokens on building slop to get rid of you.
Goes to show how fake this industry has become when VC dollars have flooded it.
Somehow it is fine to vibe code infrastructure or security because someone (with a clear vested interest) wants you to spend more tokens at their casino because that is how they "win" at the casino (which they work at).
Except in reality, this part of software is critical and irresponsible to
'write loops" and we all know that he doesn't believe what he is saying.
It's very very clear they're eating their own dog food, in a product space built on tech that didn't really exist publicly 5 years ago, to the success of billions, that people increasingly depend on. Maybe I'm an optimist, but I can't fathom the intense negativity or perspective of failure here.
Don't use it. Maybe wait a few more years. If it's not valuable/useful, then not using it, while everything matures, will not be a problem.
Meh, this is the "must be the veganism" fallacy: if someone knows you're vegan, then any ailment you might have, no matter how ubiquitous in the population, must be somehow due to your vegan diet and no more details are required.
Except now it's the "AI did it" fallacy where if you know a company uses AI, even infra scaling issues must be due to AI, and if you had just used less or no AI, you would have been spared even though that has never been true.
The usual response to this goes something like "well they made claims that AI is good" therefore anything short of perfection supposedly debunks the claim.
This is literally they saying they are letting their LLM run wild(ish) and seeing the status.claude.com we can see the result.
This is a case where the outcome is the direct result of the engineering practices like the ones they describe.
PS: Yes I use Claude, Coded, Amp and Cursor agents every day so I am not saying here LLMs are not valuable.
LE: They did not made claims that "AI is good" they made claims that developers/computer engineers are not needed anymore in the near future. Thats is a stronger claim and has a direct relation with a product they have which needs computer engineering (yes infra counts too) and which seems to be down more than we expect as a good quality bar.
You just said "it's not the 'must be veganism' thing, it's the 'must be veganism thing'"
Unless you have inside knowledge of their infra ops and management tools, it is just guessing and blaming veganism. For all we know it could be tools from Nvidia or anyone else failing under massive load.
It could be the veganism. Some things are. Leaping to it as the only possible explanation for every ailment is exactly the fallacy.
No. We dont need metaphors like that with veganism (which touches ideologies also) when talking about engineering and a company that promotes out loud that engineering is done.
I have not stated anything. I just replied to a metaphor which is not needed cause here we talk about engineering problems handled by engineers in a tech company. I give you something else where this line of thought could be wrong: culture beats (and destroys) engineering practices unless regulated by law.
In this case yes this is not because of LLMs but because of company culture.
Still hard to know where the line draws because Anthropic talks about solving computer science for good as in humans need not apply.
If you're really committed to the "no difference between datacenter hardware engineering and claude code harness engineering, they must all use the same practices, anything true for one is true for the other" bit, fair enough. It seems fairly ideological to me.
But it is like that. You have zero insight into the infrastructure issue. And the person quoted above is a Claude Code developer. So because this guy uses Claude generously to build Claude Code, then Anthropic's API scaling issues must necessarily be caused by his agent loops even though scaling issues plague every tech company, no less often pre-AI.
The issue is that it's a thought-terminating cliche, and it would be nice to have one place on the internet that isn't just who can post one the fastest with the most glee to the giddy seal-clapping of the audience.
Engineering practices or best practices are much more than writing code.
So not sure what we are debating here: I see first hand companies jumping full on using LLM for _everything_ for the last 6 months (of course Anthropic longer) and without guardrails and good engineering practices the number of incidents, downtime is increasing.
Look at status.claude.com - Anthropic could at any point come out and say all those are due to third party providers.
I am also not saying here Anthropic is worse than other scaleups. But they do something different: they come in front of us and tell us they have better engineering practices.
> Anthropic could at any point come out and say all those are due to third party providers.
Why can't it be simply the case that Anthropic is struggling by their own accord? Infra scaling isn't a solved problem, much less with new, complicated, ever-changing, stateful LLM requests.
Pretty much every API-service-centric company I've worked at was in some constant state of either triaging or thinking about infrastructure health, often due to the familiar cascading problems of a necessarily distributed system.
But now with the AI scapegoat, we rewrite history to pretend us humans solved infra scaling, so any issues today must be caused by AI and any related superstitions we want to tack on.
Another data point: GitHub is extremely insistent its employees maximally use AI for internal development [0], and we’ve concomitantly seen its reliability fall off a cliff in the last year or so.
>GitHub is extremely insistent its employees maximally use AI for internal development
Or it could be that GitHub saw a 14x increase in commit volume last year[0], and we've concomitantly seen its reliability fall of a cliff in the last year or so. Given that Microsoft is leasing additional space on AWS(!)[1] to handle the additional commit volume, my personal money is on commit volume growth being a bigger issue than internal use of AI.
Internal use of AI may have been an issue. Commit volume growth may have been an issue. Unless one has direct knowledge of their infrastructure issues, claiming to know is quite literally making exactly the "they are vegan, their illness must be caused by their veganism" argument the GP commenter was talking about.
There’s a difference between having normal levels of difficulty and bad luck, and having people blame those on the wrong thing, vs having extraordinarily miserable quality and having people find the obvious difference. Potentially yes, they might have terrible wiring in their office or a crippling fondness for vim. But if I were their PR department I’d be talking about that if it was the problem.
If you go around bragging that you use AI for everything as part of your marketing plan, then don't be surprised that people blame you heavy AI usage when you have a problem.
"..people who followed a vegan diet had noticeably low levels of iodine in their bodies, an element that is essential for growth, bones, and brain function. In addition, vegans had lower bone health scores..." - https://www.bfr.bund.de/en/press-release/vegan-vegetarian-be...
There are a lot of nutritional blind spots in vegan diets.
It is a diet that requires exceptional planning and intentionality to be at a baseline of health similar to a balanced omnivorous diet.
So indeed, the "it must be veganism" is not an unfounded concern when health complications arise, in a very similar way to "it must be the AI" is a valid concern when software issues arise.
This isn't really the place for this, nor does it matter to my analogy.
But I was more getting at, say, staying out of the sun or being skinnyfat as a vegan, and suddenly you look "sickly"/"frail" when you'd be given the grace of looking like most people otherwise.
A similar analogy would be someone saying "well, of course you do" if you have any malady while having been vaccinated. My point being to bring up the thought terminating cliche of it compared to doing the necessary further analysis to link the malady with the suspected cause.
---
> "Vegans and vegetarians may have higher stroke risk"
It was a lump vegetarian + vegan group with a weak CI bounded at 1.02 for 3/1000 cases over a decade. The same group also had a more robust benefit of less heart disease than meat eaters. The stroke outcomes aren't replicated in other cohorts either, afaik. But the heart disease benefits are.
> "Vegans had a 43% higher risk of fractures overall compared to nonvegetarians, as well as higher risks of hip, leg, and vertebral fractures."
The study used a single baseline questionnaire for 17+ years and looked at vegans with correctable nutrition deficiencies to see +15/1000 hip fractures over 10 years. I'll grant that a poorly planned diet, especially 30 years ago with less nutritional understanding, has worse health outcomes. Just like I wouldn't use the average American's diet to lambast an omnivore diet (compared to, say, the "Mediterranean" diet).
> "vegans had lower iodine, bone health scores" (RBVD study)
On bones: p=0.02 in 72 people with 5% less QUS score in their heel bone (not DXA nor bone density tested). No body weight mediation nor data about health outcomes like fractures, osteoporosis, and no time dimension since it was just a snapshot (cross-sectional).
On iodine: It's a surrogate biomarker from a single pee test. Study didn't look at iodine-related health outcomes like thyroid dysfunction, goiter, or clinical consequences.
I just did a build in Nemesis8 (containerized agents) and Pi appears to be working fine. Opencode is a good choice too if you're interested in checking out GLM 5.2 from z.ai.
They are different models. OpenCode is trying to be a claude code/codex replacement, where-as pi is something you build yourself, kind of trying to be an emacs type thing compared to vs-code. As in emacs it is more common to write your own extensions, where as in vs-code most people just download them.
I keep butting into the question of; why opencode, when you've got codex available? Codex is open source as well, and i can't seem to picture a situation where one would want Opencode over Codex.
As far as I can tell, they tick the same boxes- but one has the support of a big boy model provider.
oh-my-pi is a bit of a cross between the two; comes with basically everything OpenCode does, but still easy to customise.
OpenCode is nice if you don't want to do a lot of research and just want to get started right away. The OpenCode Go plan for $5 a month for your first month is a great way to do this, with good models to choose from and reasonable usage limits for a beginner.
I am genuinely curious what it tells you, as "curl https//.. | sh" has long been an enormously popular approach to distribution in the open source world. Homebrew, to name just one example, advertises a similar method.
(pi.sh also documents other install methods, like `npm`, on their homepage)
If trust and security is the issue, unfortunately "better" ideas like hashpipe [1] never achieved critical mass
I really hate the `curl <url> | sh` specifically because if your connection drops at a specifically unlucky point in time you are left with a partially executed script which if you are unlucky enough may just have been executing `rm -r ~/.cache/<pkg>/download` but it stopped at `rm-r ~/`.
Is it likely? No. Can it happen? Yea.
Just make it `curl -o <file> <url> && sh <file>` and this entire problem is gone.
> I am genuinely curious what it tells you, as "curl https//.. | sh" has long been an enormously popular approach to distribution in the open source world.
It's plain horrible. You could have, for example, a compromised server serving malware but only one out of every 100 download. The only signature you rely on is TLS.
Proper package distribution are using proper signatures schemes, are decentralized, even for some offer reproducible builds (meaning you can rebuild the whole package yourself and verify your build matches), etc.
Hashpipe is an attempt at reproducing some of those guarantees. Not unlike container pining using hashes. It at least fixes the "Jack and John installed this already and I know I'm getting the same version as they did".
Proper software distribution is signed, reproducible and ideally also uses some proof-of-existence for the hashes.
My bet is this: in the face of the countless supply chain attacks, we'll see more and more people getting very serious about security, including the security of software distribution. And curl bash'ing won't be part of it.
Package managers: ecosystem is fragmented, requiring a long list of distro- and package-manager-specific instructions. Many scripts already install through package managers, they simply make the user’s life easier.
Flatpaks: These are clearly designed for desktop applications, with CLIs treated as an afterthought. They may be the best long-term hope, but today they are definitely not as convenient or widely available as a simple script.
If you care about adoption, `curl | sh` is the only real option today, which is why virtually all project show it as the first option.
The ideas aren't mutually exclusive, and I've never seen an open source project support "curl | sh" without also supporting those methods.
Indeed, plenty of these scripts often act as a "what OS and packager do we have" mux. Just look at the source of this one, for example.
When you support an open source project at scale and/or with less savvy users, you come to see the benefit of "here, just f'ing slam this into your shell and we'll figure it out" installers. I know I have.
There is no threat model that doesn't also apply to pretty much every other distribution method.
It's just people who have internalized "don't paste commands from the Internet into your terminal" and aren't thinking about exactly what makes pasting commands from the Internet into your terminal dangerous, and how that applies to this specific case.
Further - what the flicking fuck do you think an installer is going to do on your system? Not run any commands? Because I've written installers for every platform... they ALL can run commands.
So what exactly is the complaint in this comment? If you want to go read the install script - knock yourself out (or hell, point your agent at it...).
it tells you they're just like basically every other CLI targeting project for the last 15 years? I mean is it a big security hole we all accept, yes, it is. But it's not really indicative of much. That's also how I install rust.
Incredible how we can claim productivity increases when its either Claude or Github shitting the bed every other day. It must even itself out to a net neutral gain in the long term.
We are back to the baseline. The availability of our tools isn't adding anything in the long term because the productivity increase we get from the tooling is negated by the time we're back to doing it the old fashioned way due to downtime, so there is no claimed productivity increase espoused by the pontificators of the tooling.
The bunch of MD files in the codebase is becoming "tech" debt. It's just English prose, sure, but thousands of lines of English prose. Terse. Succinct. Difficult (if not impossible) to maintain manually without LLMs. That's not "baseline"
Both articles use 2017 as the turning point date. TTS is a lot older than that. It's not difficult to find data to fit the desired point if you choose a narrow enough time range. Or location selectivity - both of those are just about the United States.
When that downtime happens is way more important than the amount of it. Imagine if your payroll system was down for 8 hours a month, but it just so happened to be the day payroll do their calculations?
Totally. The uptime metrics are deceiving imo. A more useful measure for a productivity tool like Claude Code is uptime during work hours for a given time zone. I strongly suspect at least for the three US time zones, we would be looking at a single nine of uptime for that measure.
Business/Office productivity tools can be productive at that rate. Core systems like ERP or arguably CRM can't, but MS Teams is probably already that low, Figma, Canva and several others could absolutely afford to be one nine before it affects their churn materially. I suspect OpenAI and Anthropic make most of their profit on business use cases rather than dev use cases (likely higher revenue but less profit) so this may be what sets the standard of uptime.
Works out at even more days when you consider working hours. And these downtime events never happen when I'm sleeping, always smack in the middle of the afternoon when I'm working.
I have been developing software since the late 80s, mostly CAM software for metal cutting machines, and I have been refereeing tabletop roleplaying games like Dungeons & Dragons since the late 70s.
I get the power of LLMs, and I do find them useful. But I find them useful in much the same way I find a really good set of random tables useful, or a good set of rules for procedurally generating something like a star sector for a science fiction campaign.
For my day job developing software, and for the RPG campaigns and books I run and publish today, LLMs are, in many cases, random tables on steroids. After using them for two years, even with all their improvements, I am continually reminded by the results I get that, at the heart of it, I am still dealing with what amounts to randomly generated content.
Yes, I know it is more accurate to call the process probabilistic rather than random. And yes, somebody can construct a technically deterministic setup with fixed weights, fixed seeds, fixed sampling parameters, and a frozen runtime environment. But that is like saying you can recreate a rainstorm if you get a thousand butterflies to flap their wings in exactly the right way. It may be technically true, but it is not how the technology behaves in normal day-to-day use.
For practical purposes, given the same prompt and the same apparent starting conditions, the result can differ each time you use a model. The outputs will often be highly correlated, and often useful, but they are not deterministic software in the ordinary sense.
So far, I am failing to see how the inherent probabilistic nature of the technology can be fully overcome. I understand how we got to where we are today from older neural net technology, including the systems used for vision and sound. What we have now can be very useful. But my view is that it is being badly oversold and overhyped. Its probabilistic nature is being vastly underestimated, and that is a major reason for much of the weirdness and many of the failures we keep seeing.
In tabletop roleplaying, there have been times when hobbyists relied too much on procedurally generated content and ultimately got burned by it, either through campaigns that were not as fun or products that were subpar. Each time, the lesson was the same: there is no substitute for human judgment.
Any workflow or technology incorporating LLMs has to keep humans in the loop, and not merely as rubber stamps. The human has to remain the primary decision maker.
That doesn’t disqualify humans. It highlights the difference I am talking about.
Those chemical interactions and quantum effects lead to emergent properties like judgment, experience, context, accountability, and an understanding of consequences. Those are not properties that LLMs possess, regardless of how useful their output can be.
That is not to say that, in the future, LLMs won’t be used as part of other systems that add some of those properties. But that is not what we have today, or what can be seen in the foreseeable near future.
Imagine a future where Anthropic holds your company hostage because no one can code properly anymore by hand and demands paying 200% higher price for the usage.
I would guess that they would want to at the very least 10x their prices. Remember they need to make up for training, marketing, etc.. and make a big chunk of profit on top of that to justify their trillion dollar evaluation
Developers who can code without LLMs will go extinct in couple years and there will be legends about them, you should at least have some decent open weight model as a backup
There is something to be said for how the technology stack keeps growing for businesses and what this might mean for the future.
Thirty years ago, you had an OS and you installed applications. No problem.
Later, you had to build and use apps on the internet, an infrastructure that is susceptible to DDOS attacks, government firewalls, and other security risks. Still fine, sort of.
Now, you not only have to build apps on the internet, you also have use LLMs to build apps to remain competitive with other developers. Future (human) maintainers of your code might not properly understand how it works, and if the providers of the LLMs screw up or go rogue, you are properly fucked.
There is a dependency/technology stack debt that is creating risks that need to be acknowledged.
No need to make up speculative futures based on a company only giving one model to their employees. I use Codex, Antigravity, Claude and GLM-5.2 interchangeably. Any sensible employer will do the same.
400k isn’t crazy for the FANG set but it’s still a subset of the developer market and hundreds of thousands of those jobs have been cut in the last few years as they all collectively work to lower SWE pay.
60k a year it needs to be a full irreplaceable part of the infrastructure for I think. There are very few kinds of software that meet that bar right now (certain design tools etc that have no replacement). 12k/year is in the expensive but reasonable for the right tooling category (Matlab etc.).
I don’t know what the future holds. I know the big AI companies are banking on being able to charge for a replacement SWE that works 24/7. Still not convinced these are it yet, as useful as they can be under the right circumstances.
Actual 90d uptime: 97.6838% (calculated by Codex from live data)
Computed from the page’s own data for 2026-03-26 through 2026-06-23:
- Partial outage: 43h 15m 1s
- Major outage: 6h 46m 48s
- Total affected time: 50h 1m 49s
- Major-only uptime: 99.6861%
I signed up for paid plan on Claude just 3 hours ago for the first time and was scratching my head on how that thing gets praised so much if I can't even send a question half of the time....
Perhaps I am taking this idea a bit too seriously but I imagine that it might not work because of VPN's.
but VPN's can be detected and perhaps already are by these AI companies and it can lead to the ban/restriction of an account so not many people would prefer to use VPN.
Then, theoretically speaking, I suppose that it might be possible to perhaps toggle off these AI companies for enterprises or licenses of dev's
Though I imagine that it would mean taking an ID and having a special dev tag so as to not remove the general purpose chat bots that these sites still operate.
I do imagine that it might be really interesting to have a single day where AI esp closed source is/are turned off and see how that pans out but looks like till then claude is sprinkling its downtime throughout any part of the day/month randomly with their downtimes.
It would be hilarious if they don't know how to fix it because this was built by "running loops calling Claude" and they haven't the faintest idea of the present underlying architecture.
I have two sessions going. One is fine, one keeps timing out. Both Opus 4.8 in Claude code in terminal. Must have them routed to different to different infra that isn’t equally impacted.
Protip: in the olden days we used to be able to read and write code ourselves. Worth trying while Claude is down! You might have fun and learn something!
I'm perfectly capable of programming without internet or AI but I would admit it would take longer and in the modern world we live in it's often not economical to do so. After programming for over 20 years you start to get in that flow automatically at least you used to do so. I don't know if people starting out to program will be able to, but most experienced developers will feel this way I assume.
> I'm perfectly capable of programming without internet
Me too, but let's be honest, I'm not talking about "Hello world!" experiments, I'm talking about developing usable software. I'm pretty sure, you won't be patching a Linux kernel driver on your own machine without googling stuff.
I've learned to code years before the Internet, but we've had it for so long, I'm honestly not sure anymore if I'm truly capable of building [real] stuff while offline. And I can't just ignore it, there's a feeling now, that with AI advancements, I may soon no longer be able to code efficiently without any AI.
It depends on the language but agreed. If I didn't have internet or AI access, I'd still be able to pull out manpages or dig into source code.
I wouldn't like it and it'd be slower, but I still understand my environment in sufficient depth to work without external info if I absolutely have to. Even with AI, once in a while I ask it to just give me some hints instead of solving something for me, so I'm forced to do the work.
Hey you. Touch grass. Go outside. If a minor downtime of a developer tool triggers you, it means you likely have heavy anxiety. Don’t worry about it and calm down.
Anthropic has massive capability issues due to massive user growth. It happens often when EU and US work hours collide. They have smart people working on it. Don’t waste your energy complaining.
I request an official statement from Anthropic explaining how they're going to limit outages in the future. Elevated errors almost always means its down for me and I can't be that unlucky statistically speaking. It seems that Anthropic does not have a good grip on the ops side of things.