I know many (big) places that are already back-paddling.
From token leadership boards (more = better) to now saving on AI cost, because the benefit was less than expected. As if this wasn't obvious. Sigh.
Sure MS, Meta, etc, are still laying off engineers to have money for data centers. But I am 99.9% sure that they will thoroughly regret that move in less than 2 years.
I don't get the obsession with more code. Whatever happened to good development practices? Many of which, if followed allow LLMs to work more efficiently ironically enough. I would rather hear that you've fixed bugs that have stumped your dev team for years, and have had next to no outages, to me that's more valuable than just slopping features out the door.
While I agree with your sentiment, very strongly actually, there has always been a tension between what I'll call "engineering teams" and "product teams". The engineering teams want the long-term platform stability, performance, etc. The product teams want to push that new feature out, quick! The overall business needs both (types of thinking and doing) to survive, of course, but not all leadership sees the world through the lens engineering team lens. I'd argue most actually lean the other way, which is understandable though unfortunate. I'd also posit that since these product teams are often "closer to the business", those people typically win favor with the real decision makers.
This tension is why I refer to myself as a product engineer. Teams where non-technical product managers (ie. your "product team") become feature factories that eventually drown in their own tech debt. Teams where navel-gazing engineering managers (ie. your "engineering team") have too much power end up with premature platformization, and beautiful architecture diagrams while user needs go unmet.
The key is to understand that computers exist to serve human needs, and the less code / UI / hoops to solve a problem the better, and to think end-to-end. If you are able to do that, most systems built by committee have a lot of low-hanging fruit that can be improved without much sacrifice. In practice, political realities and operational concerns may come in the way, but I firmly believe this is the only way to successfully operate as an engineer, and it will only become more important in the age of AI.
If you build your platform well, then you can execute fast, but also safe.
I push "money and data are sacred"; if we lose either of those, someone cries. Build the platform slow and fast. Governance, scaling etc sits here.
Provide reference APIs and let the product team one shot whatever they want and need, build fast and experimental, but your platform enforces constraints.
Best of both worlds (though feels like we're back in backend vs frontend development)
>If you build your platform well, then you can execute fast, but also safe.
this is extremely true, but it seems less and less like anyone wants to invest in a solid platform and just wants to jump to executing fast and unsafe. and then pretends that ai can definitely do both when the experience is quite a bit different.
I have seen at least one example develop in real time. You typically spent years building before you had a platform, then lack the enthusiasm from senior leadership to invest the next 12-18 months with no new features and hard to quantify - maybe no - returns. Then you get displaced by newcomers who are "platform native"; what this really means is their engineering misteps haven't been discovered yet, or even if they guessed correctly the world has turned and made their previously correct solution now a failing one. And now we have something called "AI-native" that is going to solve this once and for all with some sort of magic, and you can't lead these companies unless you truly believe in the magic.
Features don’t usually come from product teams. They come from above: some Head Of, VP Of, and the like. Those people only ask, they know very little about execution. They want their wishes come true in a timely manner
I'm lucky I guess, I feel like the product team is on "my side" with respect to building a resilient system. They understand that bad architecture shows up as a fragmented and friction-full UX.
The demand for new features come from sales and executive people who want to check more boxes on a customer's checklist.
One thing I've spotted several times throughout this year: the larger volume of code leads to exactly the sort of problems it always has. It just happens real fast now, with fewer people. Problems that used to take years to erupt now manifest before an MVP ships, but people just sweep it all under the rug in the name of moving fast.
The rate at which software is produced has never been higher, nor has the rate at which it is thrown away for being broken and useless. It’s not at all obvious to me that if you control for both of those rate changes that the amount of useful software with a lifespan measured in years created per developer (ie, the software that actually has long term value) has actually changed all that much.
I've spent most of the past two weeks trying to get my AI to stop writing so goddamned much code. It's great that I can, say, completely switch between two databases, but I wish the AIs would develop a concept of energy spend or something and spend maybe a moment checking whether something is already implemented in the code base before blasting out 2000 new lines defining a new database access layer, and also rewriting random code it shouldn't be touching for some task hundreds of lines at a time just because it can. I don't care how easy it is to write some tests that sort of kind of cover it hopefully maybe... no new code is as tested as the stuff that has been doing it in the field for years already.
Just because it has the power doesn't mean it should be used.
I'm not asking for everything to be some sort of beautiful code whose concision would lead Knuth to tears. I'm just asking that you not slather dozens of existing functions with three new parameters and dozens of lines of code each that basically do nothing, to cover a use case that I accidentally vaguely alluded to, if you read my prompt a certain way, but isn't really what I was interested in.
I drastically improved the reliability of an internal IoT device so it can tolerate all kinds of abuse (abrupt power loss during update, unreliable SD card, bootstrapping over network when the SD card is missing/damaged, fail-safe updates, recovery in a network isolated environment). This is the first release that I really leaned into AI, having been pretty skeptical before.
I think AI has been a net positive and has been used responsibly at my work. It can be used to make existing products better and more secure. You just won't naturally hear about that. There's some challenges like emerging social mores but overall good.
> It can be used to make existing products better and more secure. You just won't naturally hear about that.
Why? I keep asking for examples of LLM being used to make code more simple and elegant (which I would say is adjacent to making software better, if not more secure), and never get a reply, not a single one. I'm not claiming it doesn't exist, I know it must exist. If only via a lots of careful massaging of system prompts and harnesses, even if it takes too much effort to be worth it, I would like to see it.
Of course people can't share internal examples, but this is all over the world for many months now, where are the open source examples? Not "something new someone couldn't have done before", but someone making their precious even neater (while also adding things to it, but not primarily). Put differently, can LLM delete code, can they curate it well, can they not write weird comments nobody wants or needs, that sound like echoes of the user prompts and responses to them, but just explain tricky parts of the code, like a human would?
Also, IMO a VERY useful thing to learn about a tool or a chemical or whatever is, "what happens if I keep applying it to the same small thing over and over?" -- because that's the level of scientist I never evolved past, the kind of stuff you do as a kid -- , whereas a lot of this seems to be more about how it feels to apply it here and there until it looks good, and then move on and never look back. Which really isn't much at all, especially given we're such a short time in. I mean, how useful were mugs made of lead? Are people aware of the (in hindsight) insane things that were done with x-rays as those were invented, before the potential for harm was known? I bet they had a great time playing with that stuff, for a while.
What if in 5 years it turns out that a few years ago, LLM got into this habit of implementing certain things in a certain way because it seemed the best to "them" (the training, whatever), but there's a security issue with that -- what do you do? You can only let more agents loose, and hope they don't miss anything, because you cannot even do something like "patch PHP" (which is not my thought, Casey Muratori mentioned that in on of those podcast things).
Fully leaning into LLM doing as much of the coding and understanding as can be handed off to them seems insane to me, I don't mean this to offend though. It just does, like suggesting we go to Mars, because everything is better there, and to do that, we have to blow up Earth. Or saying to magically be able to fly from now on, we need to saw off the branch we're sitting on. Even if I thought life on Mars was the future or that we could fly, if we just thought of it the right way, this isn't a way forward IMO.
We used to understand that code was harder to read than it is to write. It follows, directly, that better engineering discipline means shipping LESS code. Less to support, less to understand, means a product that is easier to maintain and extend.
Now, token calculators spew out code at a pace no human can match, which completely removes the barrier of effort - which was one of the largest natural counters to shipping massive amounts of slop.
I'll be impressed by a coding model when it writes less code than a human, and spends more time refactoring than it does just adding new layers. Until then, this vibecoding is just madness.
it really isn't a flex ppl think it is when they're burning tokens churning agents that are working for "hours" doing work
we don't need to read or write code anymore, but we do need to write (as humans) the instructions for agents.
having ADRs, product specs, acceptance criteria, etc. that are properly maintained leads to agents working for just a few minutes and pumping out reliable code.
the more vague the work for agents, the more randomness and inefficiency there will be. that leads to bugs, constantly re-writing code, etc.
imo ofc, i'm not yet running a billion dollar company so maybe i'm wrong.
Can you please point me to any project based around spec-driven development with agents which has worked well long-term, ideally one that isn't another agentic devleopment tool?
none of these types of pieces of content focus on what was created for users, its always about how sky high their "shipping" metrics are (PRs merged, LoC, etc). it makes it extremely hard for me to care, as they all boil down to "im a dev who made some loops on loops on loops that writes a lot of code and spends a lot of tokens."
i will say, at least this article says "we merge 5x more and our revert rate has stayed the same." so thats at least a metric i can start to respect
Any good product leader will tell you, sometimes not building something is actually good for the customer. Slower dev cycles have you the ability to take time with decisions and make adjustments. The obsession with shipping more code is making things worse for users and I would argue, damages the company in the long run.
Eveyone has a backlog. What happens when you clear it in a month? Do you suddenly have better ideas about what you want to build? Do you suddenly have better intuitive designs? Do you have more users or the same users using the product twice as much? In most cases, the answer is no.
AI gives you the speed to execute when you need to, but knowing what to build and how will still remain a bottleneck.
Now the real question is "is the revert rate the same because you still have the same process, or do you fix-forward more often now, so your revert-rate is because of higher-severity incidents?"
I mean we (huge tech company with ~2000 devs) literally never actually revert anything, in my 2 years here I have quite literally never reverted a commit. We have a staging/release cycle that means nothing ever needs a revert, you just fix issues in followup commits on trunk
This is another one of the infinite meaningless metrics that AI cash burners love to use without context
I think the HN crowd overestimates the quality of software. There is a (very) long tail of truly abysmal software that people are forced to use all the time. The reason it gets to be abysmal is because it’s often back-office admin software that helps people file forms for, say, logistics paperwork or school district expense tracking. AI-native engineering teams can create dramatic improvements for that long tail because the people running that software are either incompetent or checked out, and a pull towards average quality by AIs would yield great improvement.
One of the most uncomfortable truths about our profession is that there is literally no floor to how shitty software can be and yet still make people billions of dollars.
Reminds me of an incident where I had to go pick up a missed delivery (an expensive photo printer) from the dispatch warehouse because the shipping company couldn’t get a signature after three attempts.
Whatever, I drove down to this set of huge industrial warehouses with trucks coming and going by the dozen and found the customer service office.
“That’ll be half an hour, I recommend waiting at the cafe.”
“Half an hour for what? Aren’t all the packages right here?”
“They are, but looking up the package ID in the system takes at least 30 minutes, maybe 33.”
“33!? Wait… don’t tell me, let me guess: it has been getting slowly and steadily worse over years? Predictably taking longer and longer, right?”
“Yeah, six months ago it was 27 minutes! How did you know?”
“Umm… isn’t this a bit of drag on your operations?”
“Absolutely! Everything takes forever, every shipment is delayed, … <long rant> …”
“I… bet. Why hasn’t anyone fixed it!? It’s a one line fix!”
“No budget!”
This was one of the biggest logistics companies in the country about to undergo a multi-billion-dollar merger.
Their core business software was missing database indexes even on primary key columns!
Usability of internal tools is just never considered. Priorities are driven by management. They invariably just don’t care about feedback from the peons that actually have to use them in their day-to-day jobs, even when it results in tremendous inefficiency.
This also allows companies like Workforce to make billions selling the most godawful software imaginable.
> There is a (very) long tail of truly abysmal software that people are forced to use all the time.
We generally call it “enterprise software,” and it’s usually abysmal because the people buying/funding it are VP-level executives who care about compliance/reporting/visibility and who never have to actually use the software.
This actually comes up on HN pretty regularly. Or used to; like many topics it’s taken a back seat to AI.
...and now with AI much more of that dreadful software can be built with all the bullshit featuritis that said managers who don't use it think they want.
You're thinking about this at the wrong level. This org has made the decision to be AI-native. They aren't A/B testing being AI-native vs not being AI-native. That's the kind of thing "the market" does; if a company's markup is higher than another's it can attract more investment or its equity can trade for higher pricing on public markets.
I don't think it makes sense to A/B test "AI-native" vs other regimes of development unless you are a large, decentralized company like an F5 company or a Big Tech. How will a team of 25 (the size of the company if you read the thread) have enough output to even A/B test the difference?
It was the article that used metrics related to engineering velocity as evidence of success. It makes sense to ask how this is impacting their business metrics. The company is not in the business of shipping more code. I don’t see how asking how this ties back to core business results is thinking about it at the wrong level.
> It was the article that used metrics related to engineering velocity as evidence of success
Success for what? The article opened with: "Last week Will Larson (@Lethain
) asked how leaders are evaluating engineering managers for AI skills. My reply
, a list of questions to think about, got a lot of bookmarks.
More broadly, I've been thinking about how to build an AI-native engineering org for a while. Sharing some lessons from the last nine months of building at FurtherAI."
The opener is "how to build an AI-native engineering org" and the context is a company of roughly 22 engineers (from a reply.) Nothing here says "evidence of success"
> I don’t see how asking how this ties back to core business results is thinking about it at the wrong level.
In an company of 22 engineers, how would you make these measurements? Would you split your company into a velocity focused group and a quality focused group and make them compete? How? What would a customer motion look like? What would these groups share? This makes no sense. This is like asking a small company to A/B test teams that used a computer vs didn't use a computer.
A larger, more mature organization has the luxury of splitting business units into AI-native vs not-AI-native workflows and comparing COGS/business objectives against each other. I know of very large tech consulting services who did exactly this around the beginning of the cloud era to figure out which aspects of their business they wanted to place on clouds and which aspects they kept on-prem, and even then that was messy because part of the value of on-prem is the scale benefit you get by owning and operating your own DC. But no 22 person engineering team I saw A/B tested this out. Companies were founded either with a cloud-first philosophy or an on-prem philosophy.
This is a 25 eng startup. Anything can go wrong here. A hike in interest rates can lead to a down round causing the company to go into cost cutting mode, leading to a decrease in product velocity and prod incidents as infra is cut to the bone, in order to cope. Your top engineer can get sick and quit and your next in line isn't strong enough to keep the ship going.
I'm curious if you've ever worked at a startup, tried to invest in a startup, or considered purchasing a startup's product to know what it's like? The whole point is you're taking a risk trying something new and different. If you're just doing what a much bigger, established company is doing, you're fighting against an incumbent who already has much lower COGS than you.
Taking a bet on being AI-native as a startup is taking a bet that use of AI will result in lower COGS than the incumbent. As the old saying goes, "Nobody got fired for buying IBM", and yet Amazon was able to crack into that space by doing something novel. If Amazon was just IBM, why not buy IBM? The ultimate test of AI-native business models will be if these companies become the Amazon of today, or whether we'll write about them the way we wrote about "fifth generation computing" in Japan.
>> I've started largely using this site in a read-only fashion. The SNR here now is too low and I'm too busy. Found my time better spent in more focused offline and online communities.
I've gone through a few recruiter screens lately and the number of people who mention "number of deploys per day" as a metric they are proud of causes my eyes to roll.
What can you possibly be doing as a product that justifies you releasing 60 times in a working day?
It's certainly better than the old models of releasing on a quarterly cadence with the test team and all that, but it feels like the metric has become the goal when you are releasing that often.
That's why I'm sceptical about the value of ai in software development. Much more can now be done but I don't see costumers paying more for it. The only way it can work is by laying off software engineers and saving money on their salaries.
Tend to agree. AI has failed to produce more business outcome. The only thing left is cost efficiency by reducing headcount. Thus 5x more productivity means needing just 1/5th your current staffing.
and IME the #1 rule for us is work in a profit center, not a cost center. When you're cutting there's a law of diminishing returns, but the floor is so incredibly low you don't want to be near it just to clear the blade.
since when does an engineering org care about revenue or customer satisfaction? Most software devs have no idea how the business they work for operates and think their direct deposit every two weeks is a fundamental law of nature. Also, many software devs are anti-customer and think customers are wrong even when the customer is talking to them one-on-one describing their needs in plain english.
This post is funny to me, because I feel it falls well short of what a proper, forward-thinking "AI-native" engineering org would actually look like. That is, the author is presenting this as kind of a radical departure from normal engineering, but I just see it as normal engineering with more AI. I don't see that as AI-native.
To me, true "AI-native" engineering will be one where questions regarding how humans will do things like review code are irrelevant. You explicitly throw away any concept of typical software engineering and, instead, differ to the AI. Your job, as a human, is solely to do what the AI can't physically do. That's not the same as closing gaps between what AI is currently capable of doing and what you'd need it to be capable of doing. It's literally enforcing that anything an AI is able to do is treated as correct in itself, and you, as a human, only support it by being a human.
If that sounds outlandish, that's because it is. We don't have such capable AI systems yet. It's possible we never will. But that's beyond the point. The reality is trying to make an "AI-native" engineering org by just layering AI on top of existing processes and systems is futile. If your expectation is that AI will continue to improve in capabilities at an accelerating rate, then these stop-gaps are pointless. You might as well just ignore AI altogether until it is able to completely take over, because that's the inevitable goal, right?
If you're willing to buy that will be possible, then my guess is that it will occur in the form of small scale startups and actual mavericks rather than someone trying to run an existing business who will simply be too risk adverse to go all in when the time comes. Some guy in his basement is more likely to leverage AI "properly" to build an actual "AI-native" engineering org because that person won't have much to lose.
> If you're willing to buy that will be possible, then my guess is that it will occur in the form of small scale startups and actual mavericks rather than someone trying to run an existing business who will simply be too risk adverse to go all in when the time comes. Some guy in his basement is more likely to leverage AI "properly" to build an actual "AI-native" engineering org because that person won't have much to lose.
FWIW I think the company writing this post is one of those small scale startups. But that being said, you're not the only one who thinks this. A decent amount of VCs and funds think that only smaller, newer companies have the potential to be truly AI native because there's too much organizational inertia in an existing business in keeping its existing human processes and so only a newer company can achieve the true "scaling potential" or using AI to scale without human labor.
I think this is all fair, but somewhat just semantics.
Most engineers are working in "normal" engineering orgs, but can feel the ground underneath them shifting, and companies believe that if they are caught flat footed, they will be out competed in the marketplace, so they are all trying to make sure they are adopting it at least in a way that lets them keep pace if not accelerate revenue.
If you are not impacted by economic incentives, you can continue to not care about this transitional state, but most people do.
Are there any companies outside of the "AI shovels" business that make more money because of AI? Every time I ask someone why they went so hard on adopting AI, I hear the same story about the fear of losing the market share or becoming less competitive. Honest question, who is benefitting from all of this? The revenues at majority of companies seem flat at best. There is a growing concern about professionals getting deskilled, the job satisfaction is not improving, wages are stagnant. Mind-boggling that we are actively engaged and often push for this experiment that hurts our own profession. Cheering for our own demise.
I use AI compatibility as a metric when evaluating new software to add to my stack. If an AI agent isn’t able to do most of what you can do in the app by opening it, I prefer to look for other options.
Does not apply to everything though. E.G: I don’t care about AI support for my non work chats/email, since I’d rather not give it access.
Reminds me of a few things, like being a meat proxy, as well as books like Company of One or Million Dollar One Person Business, and these were published before AI so right now we will see companies with only one founder and hundreds to thousands of AI agents. There already was that drug delivery company with the one founder but apparently a lot of their revenue was fraudulent. Still, it proves the point that it is possible.
> I feel it falls well short of what a proper, forward-thinking "AI-native" engineering org would actually look like. That is, the author is presenting this as kind of a radical departure from normal engineering, but I just see it as normal engineering with more AI. I don't see that as AI-native.
A company which proclaims they are "AI-native" isn't actually interested in engineering, they are trying to ride the hype wave. Companies who care about engineering just... do engineering, and don't make a big deal out of whether they are chasing trends.
The reality here is even more idiotic: someone criticizing a company for a terrible strategy they aren’t using that wouldn’t work if they did. According to the critic!
IMO most modern engineering practices are cargo culture. I’m sure in 10 years we look back on today the same way we’d look back on the days before version control was common. We are in the emailing files around stage of development
If you follow this logic, we get to the true "AI-native" engineer which is ecstatically floating in a bath of dopamine-stimulating liquid, while AI reads their brain pattern and instantly builds everything they didn't even form a full thought about.
"If that sounds outlandish, that's because it is. We don't have such capable AI systems yet."
It this point, a whole lot of teams have been leaning into AI-first, not-entirely-reviewed code since around March this year.
That means we are already 6-8 months in, and not far from having a full year of experience of what that actually looks like, albeit weirdly influenced by the increase in model capabilities during that period.
I'm looking forward to those retrospectives. Did this turn out to work OK for the teams doing it? What were the biggest failings and/or surprises? With hindsight, did this produce good quality software?
I think the biggest shift is in predictability. Will something be done today, next week, or next month - nobody knows. Will we converge on a bug free state? not sure.
So far AI works great at fast moving work which you will probably recode or throw away in a month or two anyway (e.g. a startup) but is weak at maintaining a large code base that you rely on. At least part of this is cultural, as the cost of misalignments between team members can be quickly translated into broken software.
Software robustness, at least in terms of "has the app crashed on any client machines today", is achievable. It's just expensive enough to achieve that most people (rightly or not) don't prioritize it, and the more moving parts (like third party libraries or language runtimes) your application has the harder it is to achieve. Maybe AI will lower the cost of achieving that.
When I worked on Guild Wars every server or user crash went to a shared inbox and was bucketed so we could identify and act on new crashes. IIRC we maintained very, very low production crash rates and could prove that many of the crashes were due to hardware faults.
Similarly, Mozilla tries to analyze and act on every crash report they get in the field. If a webpage visited by even a small % of the customer base was causing crashes, the team worked to try and fix the crash to deliver stability. If it was a high-traffic webpage it got treated as a high-priority task - I remember Farmville breaking was basically a 3-alarm fire back when it was extremely popular.
"No bugs at all" is an even higher bar, but it probably is possible for simpler applications with the use of stuff like theorem provers. I'd love to see someone try it in production but I think the return on investment for that is probably super super negative.
I operate on a "no known bugs" basis with quite a bit of effort invested to try to find new, unknown bugs.
The idea that every crash isn't acted on is kind of a crazy one, to me. Not every crash is actually actionable in e.g. a mobile app, but every one should at least get an eyeball in my opinion.
This, to me anyway, is the bare minimum of "caring about software quality". If you do this regularly you'll have very solid code, and it's interesting that I have the perception at least that Guild Wars was a quality product.
There's another layer, where you do that for subthreshold issues like UI inconvenience, etc, but that's bonus, I think.
For all the talk about AI making teams super productive, I don't see much talk about AI improving software quality.
It's a shame, though, because I believe there is a lot of potential there.
I mean, quality is easier with the right tooling (monitoring, linters, property-based tests, theorem prover...). But, given limited resources, it can feel difficult to prioritize investing in quality when business is pressing devs to ship, ship, ship.
But now the cost of writing custom tooling has changed. These custom tools might be sloppy because vibe-coded but we wouldn't have them otherwise! Or we might finally have decided to make that investment after too much time suffering from their absence. Now the cost/benefit changed, so why not invest now?
Have been at it since Nov 23rd 2025 with a first commit with Opus 4.5 - it was rough at first but after about year well worth it. You can also tell where so many just get stuck in the mud due to lack of direct experience with all the engineering silos (from bare metal sys admin to QA eng silos outside of standard software engineering core) it takes to roll out a new kind of technology company while scaling a new kind of commercialization model, also managed by AI and AI native from the ground up.
I was just looking at some user-facing error messages Claude generated for me which contained unnecessary/irrelevant technical details while failed to provide clear instructions or fix.
Earlier, I reviewed code that does way more than specified scope, changes public API and introduces new behavior. I had to almost redo the whole thing.
I thought to myself, no way I am going to ship the code as written without actually reading through it (or as some people put it, only looking at the "intent"). I'll get more complaints from users while having absolutely no idea what's going on with the code or actual behavior.
It's a puzzle to me how these orgs can pull it off.
It really depends on the kind of product you're working on. A lot of developers (and managers and business leaders) assume that their workflows will apply to other kinds of products; and are often very wrong.
For example: A few days ago I saw a quote that over 90% of current Y-Combinator startups vibecode their products. The problem is the author misunderstands context: Startups sacrifice quality for time to market. In other cases, if you're building something that isn't technically challenging, (because it's 80% similar to many other products out there,) you can probably vibecode it.
In contrast, if you're doing something technically challenging, it's to your benefit to be a lot more hands-on. I recently started getting excellent reviews from copilot of my hand-written code. (Or, I should say, the hand-written parts of an AI-assisted pull request.)
I think it also depends on how well defined requirements are. I can have Claude nearly one shot a Terraform provider but it's just taking 2 API contracts and combining them together (the provider contract and the API you're automating)
On the other hand, some things are incredibly product intensive and poorly defined and it seems LLMs will latch on to an arbitrary direction with much lower success rate.
It is also worse in "brown field" where there's lots of tribal knowledge and technical debt. Two systems may both exist that look very similar but oftentimes there's the one on life support that almost no one uses or cares about and the LLM doesn't know it should largely ignore that until it's thoroughly documented (which can be a significant undertaking)
The people / orgs that can actually codify their strong software engineering principles and practices into "a harness"[1] are the people / orgs that find success with these tools.
Even then, the actual development of a harness for interesting problems is a difficult task. It requires more skill & experience than people realize, and gets harder as you scale things up.
I remember optimizing a huge build that took days. We optimised it to only hours on a cluster - it took years to get this right. We realised at the end of this that the packaging system (handled by another team) was taking about 2x longer than the build because there was one particularly huge zipfile that sat there compressing for ever and ever and ever while the rest of the cluster was idle.
As in this story we blew enormous effort optimizing something that was very difficult and because of our blinkers (and our management's blinkers) we missed that it was only able to achieve a limited result. We didn't look at the whole system.
Spent few minutes reading the infomercial -- regretted it. Engineering principles form the bedrock first, and AI-tools are made to fit in them, not the other way around as this infomercial propounds. Its linkedin vibe is a big turnoff too.
It describes a team spending 30x payroll on tokens - $3,000,000 in token spend per month - building a multiplayer game where incidents and bug reports from users are passed directly to Fable 5/5.1 agents who then act on them and deploy the changes to production.
"We consider it an outage if we don't ship every hour"
It's worth reading because it's either a complete disaster or a glimpse of the future, or quite possibly both.
well, i did go check out the company spawn from the article. its certainly interesting how hands off their process is and how much they ship, but holy shit the games were fucking horrible. there was a direct WoW ripoff (right down to the assets) that was unplayable, a racing simulator where my controls stopped working 10 second in, and some sort of RPG that i started and just never loaded. so i will take what they are saying with the heaviest possible grain of salt.
Yes exactly this is my standard for everything now. You can spend an unbelievable amount of tokens doing all kinds of things that are interesting as processes in their own right.
at the end of it: is anything of value delivered to customers? the company might say "well users make the games with AI, we just serve them and host the multiplayer." okay, but again, what value was created? they brag about being one of anthropic's top customers; congrats you are handing VC money to anthropic. why should i care? theres no answer for why anyone should care, because to them the system and clout is the product, the product doesnt actually matter.
at least thats my read on it. i dont know how else to understand whats going on here.
That actually sounds really cool. Assuming it has access to a well-tuned predictive analytics mechanism that allows it to derive that changing the game in this way would increase satisfaction and revenue from you and other players like you, sufficiently to offset development costs and dissatisfaction to others, then it really should.
I could definitely see potential in an async co-op game based on this idea. A sim harness, some pbt to avoid broken states, and something like a modified nomic ruleset might get pretty far.
This is amazing and somewhat insane. It's more like they are building an organism with an immune system than a code base. Interesting that they say only Fable works for them.
I’m building the same but with a malleable SaaS. Basically when you sign up, you get the software deployed and an agent that can control that deployment and code.
But the process is slow. The user wants to add a feature to their app so they ask the AI. The AI has access to a terminal where it can clone, fix, and push. This takes time and isn’t instant.
Once AI can abstract this part, software will change forever. It’s what this post alludes to, does code, PR, etc need to reinvented for this?
Once so, no one would ever use cookie cutter software because you could build your own and maintain it using just an agent.
Yes but now that cookie cutter software can be the starting point. I’m not saying to build it from scratch. There’s probably an open source alternatives to the software you are trying to replace. Use that directly or tell AI to clone it and build it in a stack you understand.
Before AI, the businesses that used Zapier and automation were able to pull ahead of businesses that were still doing things manually. A business that has a custom built software with an agent that can change it on the go will provide even a greater advantage.
Oh it's Jacob - I used to work w/ him on his acquired startup, Pluto. He was very AI-pilled; I'm not surprised at the results. I remember wading through the AI-generated Python he had written. I think Pluto had something like 500k or 700k lines after 18mo. Everything was duplicated; there was no reuse of anything. Dozens of features, which mostly worked, not all did.
As a product that I would regularly use, Pluto missed the mark for me. As a SWE in 2023, I couldn't meet Jacob's mentality. Today, it's how a lot of us ship software.
I didn't love the product but I have to respect the process: he was able to sell his company. For some, what he did is success.
Was thinking of that exact article while I was reading this, we must be following the same folks. Highly recommend folks reading it, even if it feels insane in some regards it's a fascinating look at what the limits of this stuff could be and I found myself with a lot of new ideas to consider after reading it. Interesting times.
Edit - after writing "we must be following the same folks" I realized that you are, in fact, Simon Willison - I saw this article because you posted about it on X, haha.
This is a complete and utter disaster. They are making spam games. Like miniclips but for spam. A semi functional website to make what is basically youtube playables.
And it's devouring tens of millions. I know what tens of millions can make in the video game industry, it can keep a small team of good devs and designers and artists making something special that's _really_ enjoyed by tens of thousands of people.
We don't want more games, we want more great games. This will never, ever, make great games.
Look on ye mighty, and despair. I really don't care how this sausage was made because the only thing you need to know is that it shouldn't ever have been made.
As a (light) gamer for two decades who has been reading about the gaming industry for the past few months, I cannot understate how hard it is to create a good game, let alone a great game. So much money was spent/wasted on hopeless endeavors.
This works as we've been doing it since Opus 4.5 (Nov 2025) with every user request, feature and bug. It's the future without a doubt. The feedback loops evolve and get stronger depending on how well you fine-tune or train your system to optimize on its own.
That’s the dumbest thing I’ve ever read. Spaffing millions per month on this slop is the kind of thing people suffering from AI psychosis do. There is no way you need “thousands of agents” to build any product.
The more agents you throw at something, the more they collide. Those collisions generate more slop, which requires further slop processing agents.
> A default that made Savi think longer on every message, landed at 10:22pm and reverted by 10:40pm: that's not right, undo it it's too slow and not jazz.
I giggled on that one. It feels like children playing having a job. If they pay that much for tokens they have to have at least 80k customers paying 50$ a month. Seems huge for something like that. Or they have a lot of investors. Anyway does not feel like something going anywhere.
This from that tweet stood out to me: "In general, we've seen that OpenAI's models game their own checks"
In the past when working on agentic stuff I've observed models gaming checks aggressively, to the point that when we were validating log files to see what the agent had done, the agent responded by editing the log files to pass validation. I can only imagine how bad that kind of behavior is at scale, and it sounds like they ran into it.
I see a lot of builders (myself included) falling into the trap of focusing on the wrong things. Development velocity doesn't matter per se, what matters are things like growth in profit or traction. It's really easy to conflate automation/systems/process with actual business results. Movement ≠ Progress.
Working at an "AI native" company and these questions and stats are far too hand-wavvy for me.
What works:
- Empowering engineers. Giving engineers more trust, ownership and product decision-making leads to more engaged engineers vs. micromanaged tickets. Engineers who are treated like owners tend to notice cross-cutting problems and push to get them cleaned up.
- Hire for diversity. Not race or gender, but diversity of thought. Different ages, backgrounds, approaches to AI, focus on frontend/backend, etc. Race/gender diversity tends to come along for free. This is a huge edge because every LLM is spinning up the same next token prediction, so approaching problems from different perspectives gives more robust coverage across your codebase.
- Flat org structure, low ego. Interns and 20 year vets work side-by-side and teach each other things - hire for openmindedness.
- Communicate about product. What should it do? How do we want a feature to work, and how to handle certain edge cases? If the product is well understood, the code follows. In-person/hybrid is valuable.
- Architecture matters more than ever, but it is easier than ever to fix. Listen when engineers mention tech debt. Notice when a simple change touches 27 files. Senior engineers shine for this, because they tend to have a better nose for code smell.
- Reviewing still matters. Anthropic may throw tokens at reviewing, but it is a great way to share ambient product/code knowledge across the team. 200 engineers throwing code against the wall is chaos and leads to YOLO codebases. You can move fast AND keep a rough mental model of the codebase intact.
Won't there also be a point were you're shipping so fast that that in itself deteriorates UX? What if a page literally changes every time I reload? What if it changes mid-checkout process and the deal I agreed to 5 seconds ago does not exist anymore?!
I have come to realize that code (and I think possibly productivity in general) being expensive was always a feature. I am sure that a few great ideas did get lost to obscurity due to lack of access to resources to build them out. However, we now have an absolute deluge of complete and utter shite.
Agreed, for the individual LLMs can be powerful to let less technical folks build and customize their software - at an org level I reckon we'll start seeing more sinking ships as that bottleneck was actually helpful at limiting the flow of lackluster features to prod.
> None of this was free. Through the summer, much of our engineering effort went into tests and tooling, and customer-facing changes per engineer stayed between 17 and 30 a month
This is the reality I think for a lot of orgs and is absorbing a lot of the productivity gains. The gap between vibe coding a green field app from nothing which is inherently "AI native" and implementing AI tooling into complex orgs with large amounts of legacy infrastructure, integrations, ingrained practices etc is enormous. For us, it has been a big effort to enable agentic coding over a significant footprint of our codebase. And all this is happening while the plane keeps flying, with near zero tolerance for defects in critical systems. There is a lot more to it than you will understand if you just look at twitter or Hacker News.
We shipped more code!
We merged 5x faster!
We spent more tokens!
>> Every time we sped up one step, work piled up at the next... we started going through the entire process of shipping software, one step at a time...
I feel like I'm being gaslit here. You can become CTO with this lack of business fundamentals, and a basic misunderstanding of systems engineering?
For the argument lets say that this approach works and results in high quality apps. But will be users happy that app is updated 3 times a week and all updates add major changes? Because code generation is so cheap and fast it means we should even more thing about what to build to not overwhelm our users and actually solve their problems.
First you take your Blendtech and "will it blend" your engineers, then upload the result to the cloud inside what is called an S3 bucket (don't forget to set access permissions on your buckets kids!) to be digested by the clankers, sit back and profit!
Dark humor aside I think disconnecting a bunch of problem solvers from their computing equipment could really enhance the world, putting that analytical power into the world in person could move a lot of cogs, rather then marshalling JSON.
Can we ban twitter links? I have to use a hack to even read it and it’s always just slop. If you can’t be bothered to write an actual article, why should we bother to read it.
Here we are again, thirty years after I was taught the "waterfall" paradigm at business school.
When asked about "customer satisfaction" we were told, awkwardly, that became "political"
The paradigm did not survive its encounter with reality, and I do not think this one will either
The "only" 175% increase in productivity is very impressive, and seems realistic to me, when you remove the chores of writing code (thank you to coding agents)
What you build is still very dependant on what you need, and you do not know that until you build it
It is a paradox only for the "waterfall" paradigm. We rejected it, in the 1990s for a reason
Sure MS, Meta, etc, are still laying off engineers to have money for data centers. But I am 99.9% sure that they will thoroughly regret that move in less than 2 years.
Only three metrics went up: LOC, token spend and presumably release speed/frequency.
What's the markup (profit/COGS)? is it up?
Customer satisfaction? Is NPS moving one way or another?
These new tools only exacerbate this issue.
The key is to understand that computers exist to serve human needs, and the less code / UI / hoops to solve a problem the better, and to think end-to-end. If you are able to do that, most systems built by committee have a lot of low-hanging fruit that can be improved without much sacrifice. In practice, political realities and operational concerns may come in the way, but I firmly believe this is the only way to successfully operate as an engineer, and it will only become more important in the age of AI.
I push "money and data are sacred"; if we lose either of those, someone cries. Build the platform slow and fast. Governance, scaling etc sits here.
Provide reference APIs and let the product team one shot whatever they want and need, build fast and experimental, but your platform enforces constraints.
Best of both worlds (though feels like we're back in backend vs frontend development)
this is extremely true, but it seems less and less like anyone wants to invest in a solid platform and just wants to jump to executing fast and unsafe. and then pretends that ai can definitely do both when the experience is quite a bit different.
The demand for new features come from sales and executive people who want to check more boxes on a customer's checklist.
Just because it has the power doesn't mean it should be used.
I'm not asking for everything to be some sort of beautiful code whose concision would lead Knuth to tears. I'm just asking that you not slather dozens of existing functions with three new parameters and dozens of lines of code each that basically do nothing, to cover a use case that I accidentally vaguely alluded to, if you read my prompt a certain way, but isn't really what I was interested in.
I think AI has been a net positive and has been used responsibly at my work. It can be used to make existing products better and more secure. You just won't naturally hear about that. There's some challenges like emerging social mores but overall good.
Why? I keep asking for examples of LLM being used to make code more simple and elegant (which I would say is adjacent to making software better, if not more secure), and never get a reply, not a single one. I'm not claiming it doesn't exist, I know it must exist. If only via a lots of careful massaging of system prompts and harnesses, even if it takes too much effort to be worth it, I would like to see it.
Of course people can't share internal examples, but this is all over the world for many months now, where are the open source examples? Not "something new someone couldn't have done before", but someone making their precious even neater (while also adding things to it, but not primarily). Put differently, can LLM delete code, can they curate it well, can they not write weird comments nobody wants or needs, that sound like echoes of the user prompts and responses to them, but just explain tricky parts of the code, like a human would?
Also, IMO a VERY useful thing to learn about a tool or a chemical or whatever is, "what happens if I keep applying it to the same small thing over and over?" -- because that's the level of scientist I never evolved past, the kind of stuff you do as a kid -- , whereas a lot of this seems to be more about how it feels to apply it here and there until it looks good, and then move on and never look back. Which really isn't much at all, especially given we're such a short time in. I mean, how useful were mugs made of lead? Are people aware of the (in hindsight) insane things that were done with x-rays as those were invented, before the potential for harm was known? I bet they had a great time playing with that stuff, for a while.
What if in 5 years it turns out that a few years ago, LLM got into this habit of implementing certain things in a certain way because it seemed the best to "them" (the training, whatever), but there's a security issue with that -- what do you do? You can only let more agents loose, and hope they don't miss anything, because you cannot even do something like "patch PHP" (which is not my thought, Casey Muratori mentioned that in on of those podcast things).
Fully leaning into LLM doing as much of the coding and understanding as can be handed off to them seems insane to me, I don't mean this to offend though. It just does, like suggesting we go to Mars, because everything is better there, and to do that, we have to blow up Earth. Or saying to magically be able to fly from now on, we need to saw off the branch we're sitting on. Even if I thought life on Mars was the future or that we could fly, if we just thought of it the right way, this isn't a way forward IMO.
Now, token calculators spew out code at a pace no human can match, which completely removes the barrier of effort - which was one of the largest natural counters to shipping massive amounts of slop.
I'll be impressed by a coding model when it writes less code than a human, and spends more time refactoring than it does just adding new layers. Until then, this vibecoding is just madness.
it really isn't a flex ppl think it is when they're burning tokens churning agents that are working for "hours" doing work
we don't need to read or write code anymore, but we do need to write (as humans) the instructions for agents.
having ADRs, product specs, acceptance criteria, etc. that are properly maintained leads to agents working for just a few minutes and pumping out reliable code.
the more vague the work for agents, the more randomness and inefficiency there will be. that leads to bugs, constantly re-writing code, etc.
imo ofc, i'm not yet running a billion dollar company so maybe i'm wrong.
i will say, at least this article says "we merge 5x more and our revert rate has stayed the same." so thats at least a metric i can start to respect
Eveyone has a backlog. What happens when you clear it in a month? Do you suddenly have better ideas about what you want to build? Do you suddenly have better intuitive designs? Do you have more users or the same users using the product twice as much? In most cases, the answer is no.
AI gives you the speed to execute when you need to, but knowing what to build and how will still remain a bottleneck.
1) You can merge a "new change" to fix a merge, instead of reverting
2) It's unclear whether the substance of their individual PRs are the same. Never thought I'd want an LOC metric.
3) Really, I'm going to double down on the idea that merges could be fixing merges. Also happens to increase the merge metric.
This is another one of the infinite meaningless metrics that AI cash burners love to use without context
What’s the point of doing 5x more?
Whatever, I drove down to this set of huge industrial warehouses with trucks coming and going by the dozen and found the customer service office.
“That’ll be half an hour, I recommend waiting at the cafe.”
“Half an hour for what? Aren’t all the packages right here?”
“They are, but looking up the package ID in the system takes at least 30 minutes, maybe 33.”
“33!? Wait… don’t tell me, let me guess: it has been getting slowly and steadily worse over years? Predictably taking longer and longer, right?”
“Yeah, six months ago it was 27 minutes! How did you know?”
“Umm… isn’t this a bit of drag on your operations?”
“Absolutely! Everything takes forever, every shipment is delayed, … <long rant> …”
“I… bet. Why hasn’t anyone fixed it!? It’s a one line fix!”
“No budget!”
This was one of the biggest logistics companies in the country about to undergo a multi-billion-dollar merger.
Their core business software was missing database indexes even on primary key columns!
Madness.
This also allows companies like Workforce to make billions selling the most godawful software imaginable.
We generally call it “enterprise software,” and it’s usually abysmal because the people buying/funding it are VP-level executives who care about compliance/reporting/visibility and who never have to actually use the software.
This actually comes up on HN pretty regularly. Or used to; like many topics it’s taken a back seat to AI.
I don't think it makes sense to A/B test "AI-native" vs other regimes of development unless you are a large, decentralized company like an F5 company or a Big Tech. How will a team of 25 (the size of the company if you read the thread) have enough output to even A/B test the difference?
Success for what? The article opened with: "Last week Will Larson (@Lethain ) asked how leaders are evaluating engineering managers for AI skills. My reply , a list of questions to think about, got a lot of bookmarks. More broadly, I've been thinking about how to build an AI-native engineering org for a while. Sharing some lessons from the last nine months of building at FurtherAI."
The opener is "how to build an AI-native engineering org" and the context is a company of roughly 22 engineers (from a reply.) Nothing here says "evidence of success"
> I don’t see how asking how this ties back to core business results is thinking about it at the wrong level.
In an company of 22 engineers, how would you make these measurements? Would you split your company into a velocity focused group and a quality focused group and make them compete? How? What would a customer motion look like? What would these groups share? This makes no sense. This is like asking a small company to A/B test teams that used a computer vs didn't use a computer.
A larger, more mature organization has the luxury of splitting business units into AI-native vs not-AI-native workflows and comparing COGS/business objectives against each other. I know of very large tech consulting services who did exactly this around the beginning of the cloud era to figure out which aspects of their business they wanted to place on clouds and which aspects they kept on-prem, and even then that was messy because part of the value of on-prem is the scale benefit you get by owning and operating your own DC. But no 22 person engineering team I saw A/B tested this out. Companies were founded either with a cloud-first philosophy or an on-prem philosophy.
I'm curious if you've ever worked at a startup, tried to invest in a startup, or considered purchasing a startup's product to know what it's like? The whole point is you're taking a risk trying something new and different. If you're just doing what a much bigger, established company is doing, you're fighting against an incumbent who already has much lower COGS than you.
Taking a bet on being AI-native as a startup is taking a bet that use of AI will result in lower COGS than the incumbent. As the old saying goes, "Nobody got fired for buying IBM", and yet Amazon was able to crack into that space by doing something novel. If Amazon was just IBM, why not buy IBM? The ultimate test of AI-native business models will be if these companies become the Amazon of today, or whether we'll write about them the way we wrote about "fifth generation computing" in Japan.
How's that going for you?
What can you possibly be doing as a product that justifies you releasing 60 times in a working day?
It's certainly better than the old models of releasing on a quarterly cadence with the test team and all that, but it feels like the metric has become the goal when you are releasing that often.
To me, true "AI-native" engineering will be one where questions regarding how humans will do things like review code are irrelevant. You explicitly throw away any concept of typical software engineering and, instead, differ to the AI. Your job, as a human, is solely to do what the AI can't physically do. That's not the same as closing gaps between what AI is currently capable of doing and what you'd need it to be capable of doing. It's literally enforcing that anything an AI is able to do is treated as correct in itself, and you, as a human, only support it by being a human.
If that sounds outlandish, that's because it is. We don't have such capable AI systems yet. It's possible we never will. But that's beyond the point. The reality is trying to make an "AI-native" engineering org by just layering AI on top of existing processes and systems is futile. If your expectation is that AI will continue to improve in capabilities at an accelerating rate, then these stop-gaps are pointless. You might as well just ignore AI altogether until it is able to completely take over, because that's the inevitable goal, right?
If you're willing to buy that will be possible, then my guess is that it will occur in the form of small scale startups and actual mavericks rather than someone trying to run an existing business who will simply be too risk adverse to go all in when the time comes. Some guy in his basement is more likely to leverage AI "properly" to build an actual "AI-native" engineering org because that person won't have much to lose.
FWIW I think the company writing this post is one of those small scale startups. But that being said, you're not the only one who thinks this. A decent amount of VCs and funds think that only smaller, newer companies have the potential to be truly AI native because there's too much organizational inertia in an existing business in keeping its existing human processes and so only a newer company can achieve the true "scaling potential" or using AI to scale without human labor.
Most engineers are working in "normal" engineering orgs, but can feel the ground underneath them shifting, and companies believe that if they are caught flat footed, they will be out competed in the marketplace, so they are all trying to make sure they are adopting it at least in a way that lets them keep pace if not accelerate revenue.
If you are not impacted by economic incentives, you can continue to not care about this transitional state, but most people do.
Does not apply to everything though. E.G: I don’t care about AI support for my non work chats/email, since I’d rather not give it access.
A company which proclaims they are "AI-native" isn't actually interested in engineering, they are trying to ride the hype wave. Companies who care about engineering just... do engineering, and don't make a big deal out of whether they are chasing trends.
defer to the AI
"If that sounds outlandish, that's because it is. We don't have such capable AI systems yet."
That means we are already 6-8 months in, and not far from having a full year of experience of what that actually looks like, albeit weirdly influenced by the increase in model capabilities during that period.
I'm looking forward to those retrospectives. Did this turn out to work OK for the teams doing it? What were the biggest failings and/or surprises? With hindsight, did this produce good quality software?
So far AI works great at fast moving work which you will probably recode or throw away in a month or two anyway (e.g. a startup) but is weak at maintaining a large code base that you rely on. At least part of this is cultural, as the cost of misalignments between team members can be quickly translated into broken software.
> Will we converge on a bug free state?
I have never witnessed this state in my two decades of software work. Why would we think AI will magically achieve this?
When I worked on Guild Wars every server or user crash went to a shared inbox and was bucketed so we could identify and act on new crashes. IIRC we maintained very, very low production crash rates and could prove that many of the crashes were due to hardware faults.
Similarly, Mozilla tries to analyze and act on every crash report they get in the field. If a webpage visited by even a small % of the customer base was causing crashes, the team worked to try and fix the crash to deliver stability. If it was a high-traffic webpage it got treated as a high-priority task - I remember Farmville breaking was basically a 3-alarm fire back when it was extremely popular.
"No bugs at all" is an even higher bar, but it probably is possible for simpler applications with the use of stuff like theorem provers. I'd love to see someone try it in production but I think the return on investment for that is probably super super negative.
The idea that every crash isn't acted on is kind of a crazy one, to me. Not every crash is actually actionable in e.g. a mobile app, but every one should at least get an eyeball in my opinion.
This, to me anyway, is the bare minimum of "caring about software quality". If you do this regularly you'll have very solid code, and it's interesting that I have the perception at least that Guild Wars was a quality product.
There's another layer, where you do that for subthreshold issues like UI inconvenience, etc, but that's bonus, I think.
It's a shame, though, because I believe there is a lot of potential there.
I mean, quality is easier with the right tooling (monitoring, linters, property-based tests, theorem prover...). But, given limited resources, it can feel difficult to prioritize investing in quality when business is pressing devs to ship, ship, ship.
But now the cost of writing custom tooling has changed. These custom tools might be sloppy because vibe-coded but we wouldn't have them otherwise! Or we might finally have decided to make that investment after too much time suffering from their absence. Now the cost/benefit changed, so why not invest now?
Earlier, I reviewed code that does way more than specified scope, changes public API and introduces new behavior. I had to almost redo the whole thing.
I thought to myself, no way I am going to ship the code as written without actually reading through it (or as some people put it, only looking at the "intent"). I'll get more complaints from users while having absolutely no idea what's going on with the code or actual behavior.
It's a puzzle to me how these orgs can pull it off.
For example: A few days ago I saw a quote that over 90% of current Y-Combinator startups vibecode their products. The problem is the author misunderstands context: Startups sacrifice quality for time to market. In other cases, if you're building something that isn't technically challenging, (because it's 80% similar to many other products out there,) you can probably vibecode it.
In contrast, if you're doing something technically challenging, it's to your benefit to be a lot more hands-on. I recently started getting excellent reviews from copilot of my hand-written code. (Or, I should say, the hand-written parts of an AI-assisted pull request.)
On the other hand, some things are incredibly product intensive and poorly defined and it seems LLMs will latch on to an arbitrary direction with much lower success rate.
It is also worse in "brown field" where there's lots of tribal knowledge and technical debt. Two systems may both exist that look very similar but oftentimes there's the one on life support that almost no one uses or cares about and the LLM doesn't know it should largely ignore that until it's thoroughly documented (which can be a significant undertaking)
The rest are just marketers.
[1] Whatever that means to them
> no way I am going to ship the code as written without actually reading through it
Interestingly agents can help, not replace, this process
These tools will make us more productive, they will disrupt (have disrupted) our industry, but they will not replace us because they cannot
As in this story we blew enormous effort optimizing something that was very difficult and because of our blinkers (and our management's blinkers) we missed that it was only able to achieve a limited result. We didn't look at the whole system.
It describes a team spending 30x payroll on tokens - $3,000,000 in token spend per month - building a multiplayer game where incidents and bug reports from users are passed directly to Fable 5/5.1 agents who then act on them and deploy the changes to production.
"We consider it an outage if we don't ship every hour"
It's worth reading because it's either a complete disaster or a glimpse of the future, or quite possibly both.
But is the outcome good? Is the game fun?
at least thats my read on it. i dont know how else to understand whats going on here.
https://en.wikipedia.org/wiki/Nomic
In the instances where they were able to detect it. How much do you think they pour over individual traces nowadays to search for these things?
But the process is slow. The user wants to add a feature to their app so they ask the AI. The AI has access to a terminal where it can clone, fix, and push. This takes time and isn’t instant.
Once AI can abstract this part, software will change forever. It’s what this post alludes to, does code, PR, etc need to reinvented for this?
Once so, no one would ever use cookie cutter software because you could build your own and maintain it using just an agent.
One would argue that the effort to build and maintain it is why we have cookie cutter software, agentic built or not.
Before AI, the businesses that used Zapier and automation were able to pull ahead of businesses that were still doing things manually. A business that has a custom built software with an agent that can change it on the go will provide even a greater advantage.
As a product that I would regularly use, Pluto missed the mark for me. As a SWE in 2023, I couldn't meet Jacob's mentality. Today, it's how a lot of us ship software.
I didn't love the product but I have to respect the process: he was able to sell his company. For some, what he did is success.
Edit - after writing "we must be following the same folks" I realized that you are, in fact, Simon Willison - I saw this article because you posted about it on X, haha.
And it's devouring tens of millions. I know what tens of millions can make in the video game industry, it can keep a small team of good devs and designers and artists making something special that's _really_ enjoyed by tens of thousands of people.
We don't want more games, we want more great games. This will never, ever, make great games.
https://www.spawn.co
Look on ye mighty, and despair. I really don't care how this sausage was made because the only thing you need to know is that it shouldn't ever have been made.
These people have no idea what good games mean.
Do players like it?
Will the history of gaming remember this game?
The more agents you throw at something, the more they collide. Those collisions generate more slop, which requires further slop processing agents.
I giggled on that one. It feels like children playing having a job. If they pay that much for tokens they have to have at least 80k customers paying 50$ a month. Seems huge for something like that. Or they have a lot of investors. Anyway does not feel like something going anywhere.
Isn't there a point where you're shipping too fast that the product hasn't even caught on enough yet to know what to build?
I looked at some of the games and it just looks like a Roblox clone
In the past when working on agentic stuff I've observed models gaming checks aggressively, to the point that when we were validating log files to see what the agent had done, the agent responded by editing the log files to pass validation. I can only imagine how bad that kind of behavior is at scale, and it sounds like they ran into it.
I really mean this. Does it benefit the business? Does it benefit the users? When you optimize for the wrong thing, you inevitably fail.
What works: - Empowering engineers. Giving engineers more trust, ownership and product decision-making leads to more engaged engineers vs. micromanaged tickets. Engineers who are treated like owners tend to notice cross-cutting problems and push to get them cleaned up. - Hire for diversity. Not race or gender, but diversity of thought. Different ages, backgrounds, approaches to AI, focus on frontend/backend, etc. Race/gender diversity tends to come along for free. This is a huge edge because every LLM is spinning up the same next token prediction, so approaching problems from different perspectives gives more robust coverage across your codebase. - Flat org structure, low ego. Interns and 20 year vets work side-by-side and teach each other things - hire for openmindedness. - Communicate about product. What should it do? How do we want a feature to work, and how to handle certain edge cases? If the product is well understood, the code follows. In-person/hybrid is valuable. - Architecture matters more than ever, but it is easier than ever to fix. Listen when engineers mention tech debt. Notice when a simple change touches 27 files. Senior engineers shine for this, because they tend to have a better nose for code smell. - Reviewing still matters. Anthropic may throw tokens at reviewing, but it is a great way to share ambient product/code knowledge across the team. 200 engineers throwing code against the wall is chaos and leads to YOLO codebases. You can move fast AND keep a rough mental model of the codebase intact.
If that's the benchmark, developers are incentivized to produce over-engineered and bloated code.
This is the reality I think for a lot of orgs and is absorbing a lot of the productivity gains. The gap between vibe coding a green field app from nothing which is inherently "AI native" and implementing AI tooling into complex orgs with large amounts of legacy infrastructure, integrations, ingrained practices etc is enormous. For us, it has been a big effort to enable agentic coding over a significant footprint of our codebase. And all this is happening while the plane keeps flying, with near zero tolerance for defects in critical systems. There is a lot more to it than you will understand if you just look at twitter or Hacker News.
>> Every time we sped up one step, work piled up at the next... we started going through the entire process of shipping software, one step at a time...
I feel like I'm being gaslit here. You can become CTO with this lack of business fundamentals, and a basic misunderstanding of systems engineering?
First you take your Blendtech and "will it blend" your engineers, then upload the result to the cloud inside what is called an S3 bucket (don't forget to set access permissions on your buckets kids!) to be digested by the clankers, sit back and profit!
Dark humor aside I think disconnecting a bunch of problem solvers from their computing equipment could really enhance the world, putting that analytical power into the world in person could move a lot of cogs, rather then marshalling JSON.
When asked about "customer satisfaction" we were told, awkwardly, that became "political"
The paradigm did not survive its encounter with reality, and I do not think this one will either
The "only" 175% increase in productivity is very impressive, and seems realistic to me, when you remove the chores of writing code (thank you to coding agents)
What you build is still very dependant on what you need, and you do not know that until you build it
It is a paradox only for the "waterfall" paradigm. We rejected it, in the 1990s for a reason
It did not work