It will probably continue to exist because most businesses are perfectly content to outsource tech problems to someone else for a reasonable fee, and if you think that managing an agent swarm is a seamless replacement for that, I don't know what you tell you. The vast majority of business owners don't want the added complexity in their life.
And for the same reason, most SaaS will probably not migrate to some radically new UX paradigm. Some of it will be vibecoded, but the UX will probably continue to be deterministic. I think this might change once a new breed of "LLM-native" business owners takes charge, but that's going to be a slow process, and it will be hampered by prosaic concerns about interoperability, support, predictable cost, and liability for mistakes (from "the agent did a bit of tax fraud" to "the agent decided to hack my competition").
Again, barring radical superhumanity, in many cases companies do not today want their own bespoke solutions to a problem and wouldn't want it even if the developers were working for free, because the other costs to the business would still be more than they want to pay and more than another business can charge to make the problem more thoroughly (even if not completely) go away.
I suspect this is another reason you may see some companies making grabs at data and points-of-presence that otherwise make little or no sense... they're trying to colonize and defend the sources of contact with the real world before someone else gets there and locks them out.
With insignificant concerns, at times it seems a business might not even care about the cost but that's also the same areas where a cut in quality is ignored if it comes with a cut in costs as well.
A lot of SasS fall into this category, so even if AI is not great, the value of getting a solution in under 5 minutes, simply by talking to a machine, is too great a value for a business.
All those marketing sites that might've been handled by 3rd party agencies in the 80s, became in-house clip art projects in the 90s, became SaaS sites in the 2000s, and will become "CEO told his dreams to the AI, and the AI delivered while they were being driven to work" projects in 2030s.
The fact that it's bespoke or not is besides the point when writing up the requirements means the site is essentially already built and there's no need to vet vendor at all.
Now for significant concerns, SaaS will likely still have a place, but some aspects of it will be minimized. Companies focused on Wordpress, Jira, and Salesforce customizations will fall by the wayside since the main company can just say "talk to our AI agent and it'll customize it for you".
It’s exactly that kind of boring plumbing that companies want to outsource which is the exact opposite of what near term AI is good at. So yes there’s many SaaS companies that should be concerned, but suggesting all SaaS is ending anytime soon is pure hype nonsense.
It doesn't say SaaS is being replaced by agents, rather that SaaS companies will be transformed in the way described. Outsourcing will still happen, hence the discussion of headless components, and the continued demand for SaaS.
And part of the argument is that we won't just be waiting for the new breed of LLM-native companies; lots of existing businesses are being LLM-ified right now, through the process described in the article...
That said, your point can be distilled down to: comparative advantage. It still exists and so SaaS will likely still exist. But per the article, it'll probably look pretty different (and a lot less profitable) over time.
I think the reason it'll look different is that it's very hard to make powerful software that's also walk-up usable. The idea that you can provide just the core concepts of a piece of software and allow AI to interact with those concepts more directly, or allow your users to use AI to build their ideal (and ever-changing) interfaces to your concepts seems pretty convincing to me.
Of course, nothing is really preventing the SaaS side from operating the agent swarm, yes?
But to pull something from the post:
> If this is the right mental model, you should expect to see:
> An unusual amount of in-house harness building on both the build side and the sell side
Yup.
> Org structures and individual roles being reshaped around their place in the business harness
Not yet, this is the one I am most skeptical about because it's still way too easy for the harness-driven processes I've seen to go off the rails, so it's still very much human SME-driven. However, the number of butts in seats required is going down.
> AI-native startups beating incumbents in domains where the “moat” can be easily harness-ified
In most cases I've seen the "moat" is not easily harness-ified, and increasingly the money and energy seems to be going into getting to the starting line.
> All software a software company uses (on or tied to the core build or sell paths) needing to be headless so the outer harness can run it
Tech companies are definitely doing it, but I don't think this is actually realistic yet for companies whose core competence isn't software. But this feels like a matter of when, not if.
But they greatly, greatly overestimate people, that for the most part, are burnt out and have real life priorities that supercede monitoring whatever the fuck an agent swarm is. Obviously. Those people delegate to young people
The need for more compute, in the orders of many multiples in magnitude, is still needed before we even get anywhere near there. And the cost is still too high for what it is now, which is basically a science experiment after the first threshold of complexity.
There are 40,000 McDonald’s franchisees who basically just put money into a coinop machine and the machine returns with exactly the food, drinks, restaurant decor, prices, etc that it wants them to sell.
Toyota’s production system coordinates what gets made and when, detects abnormalities and directs human attention toward problems and improvement. They explicitly eliminate the need for people to continuously watch machines, while preserving human judgment
Every contract is different and the majority of the work is "the boring inefficient stuff" that can't be automated away. That is, arguing with shifty and arrogant middle management in a long series of meetings until you get some concrete requirements. Then realizing the metric shit ton of crap to untangle and test rigorously. It's not just a matter of having patience, but being unflinchingly attentive the whole time for the nuggets of gold that fall out of their disgusting mouths and then pouncing with a plan so you don't go over time and budget. AI simply isn't going to grill people like that and then execute swiftly and precisely. A lot of these jobs really are like a murder interrogation and then finding the bodies.
It's absurd that so many people believe there are unturned stones in this space. If this kind of business could be cookie-cutter, it would have happened over a decade ago and it wouldn't have made any money.
But I believe a premise of the article is that AI-ification will disrupt the power of those middle managers, probably in a few different ways.
One is that the middle managers themselves will be replaced by AI. Another is that more purchasing decisions will be made by engineers building the corporate harness on the buyer side. And, probably people will be more accustomed (i.e. forced) to talk to AIs on the other side, or at least accept that the responses are ultimately constrained by them. Going out for a round of golf with the sales guy just can't accomplish anything anymore.
Every exec team wants this transition to happen, so both buy and sell side will go through great transformation in concert.
If we're talking automation in the way that humans define a process, boundary cases, etc up front then yes you can't really automate away one-off monotony.
If AI is actually artificial intelligence, it will be able to figure that out similar to how a human would.
If you waste too much time on irrelevant topics, you lose. If you don't collect enough information, you lose. If the middle manager feels threatened, you lose. If you don't have a good production release, you lose. If there are any high severity bugs, you lose. If there are too many bugs, you lose.
The only way to win is to carve an exact path through the mess from the beginning, and that needs human experience.
I don't consider an impressive text predictor to be artificial intelligence. Maybe that means LLMs aren't AI, I honestly don't know because no one seems to care about understanding how they work or solving interpetability first.
I do expect anything that earns the banner of AI could ask good questions, weed through a bunch of word vomit to find the key nuggets, and act on them similar to or better than a human could. Anything less than that doesn't seem particularly intelligent.
I suspect for purposes of optimization and performance etc., sometime in the next several months we will start to see new popular ML architectures or variations of LLMs/VLMs that are designed to be _stateful_. So a lot of the infrastructure for managing state and memory etc. gets sucked into the model somehow.
They may end up changing or expanding the concept of an ML model to enable that.
That type of belief is what is making me anxious about AI engineering staying relevant for much longer. I just wonder if pretty soon we need to be able to build and train or customize ML models that just do everything, or at least know how to prompt an agent to do that.
In the longer term, the downstream impact is massive commoditization of software and invalidation of most existing moats. Data moats are gone if you can simulate the data with AI. Even platform effects can be sidestepped if AI replaces one side of the platform.
In addition, while right now agile startups have the advantage, at some point the balance will start tilting towards whoever has the most tokens (OR perhaps durable moats will trump even near-infinite tokens; we will have to see). Startups have a limited time window to have whatever impact in the world they are hoping to have, or to build a moat that won't be disrupted by AI, but there are few of them left in the world.
The upside is that when there is a lot of commoditization, then the consumer benefits.
- capital, as money is scarce
- network effects, as human attention is scarce
- relationships, as human attention is scarce
- research talent, assuming there exists some field(s) that AI is unable to surpass the best researchers
- proprietary data and sensors, as systems of record and action are scarce (training data, on the other hand, can maybe we simulated, but I'm pretty bearish in general on the idea of fully simulated data)
If AI levels out most other distinctions, maybe in the end we choose to give money to people we trust.
This is a hilarious premise if you work in a domain where it matters even a little bit whether the data is correct or not.
Every company can now apply the latest and greatest analysis. Data generation is where the cost is. It's where the time was spent, time that can never ever be retrieved at any cost.
AI won't solve biology, make a pathogenic virus, etc, without tons and tons of data, of both types we know and types we have not yet figured out how to generate.
Perhaps the area where AI has the most to help bio is in figuring out novel measurement technology. But it's not going to be able to reason or deep-net its way to figuring out systems for which we can't even measure the parts.
I agree this is a kind of data moat, but it's also arguably distinct enough to be its own thing.
(Edit: alternatively you just use AI to get rid of the need for data to solve a problem, like Jev did for traditional classification models)
I think current incentives definitely go against any efforts to build this. It's very hard to build this and be rewarded for it by, say, investors or your boss, because you can't really prove that your system is non-sloppy while your competitor's is (even if being non-sloppy is all that matters), because by definition your novel results are not verifiable or else the model labs will have already trained it into their model.
But the same is true for high-quality AI systems in general. In general, I think AI model advancements will make the systems easier and easier to build until some small guy accountable to no one but themselvs can build it, and then it will actually be built.
1. He’s talking about training new models and at one point, having data was valuable. Now synthetic data is being used to train models
2. Companies like SalesForce who’s moat is having all your customer data so you’ll be locked in. You could extract it but you’d have to clean it and then change it to your new schema. With LLMs, you can do that in minutes and even use SalesForces MCP or API to get all your data and leave.
It’s exactly why companies like Figma are gate keeping their MCP. They know that swapping their MCP with Paper’s or any new one is easy.
The moats are evaporating as we speak. Distribution is one of the smaller ones left, but the personal software trend might eat that too.
Being a platform for personal software is gonna be valuable, but it needs a lot of trust. (I have a nonprofit idea around this right now)
Btw, I think distribution might temporarily become less important (because with better AI you can actually pull so far ahead of competitors quality-wise and therefore succeed despite a distribution drawback), but long run it actually becomes more important because of AI persuasion and commodification? If you are the super app then, well, you are the super app
I couldn't have said this better myself. If you're not owning your harness, and you're not owning your model, then that leaves very little moat for any AI native company.
My current project is created by one-shot prompts. That’s not some kind of parlor trick. It is the framework for creating and evolving the product.
Many people laugh it off as “unserious”. But as I said, I see this happening in my place of work where people are paid lots of money.
https://jaisenmathai.com/articles/sojourn-for-ios-was-45-one...
Google isn’t successful because it has the best engineers, it has the best engineers because it is successful. Google is an extremely boring business: show adverts to make money. And they made so much money. The next Google is not going to be the company with the best “harness” it’ll be the company that has an obscenely profitable product.
Founder mode, what was considered a panacea just 12 months ago, is defined by a founder giving a shit about everything. You don’t succeed by handing everything off to an army of ~agents~ consultants.
This vision of the future is nonsense that will not pan out. You can add that to the HN AI predictions. At no point will real businesses be “harnesses” around AI models.
And on this, these are examples of tech companies filled with nerds who love novel new technology. Of course companies like Ramp and Stripe are spending huge amounts of time and money on taking this new technology to the extremes with “harnesses” and ”factories” because that’s what the nerds want to do. What the nerds want to do is not a sign of how the technology will be used in future, it’s a sign of what is most fun today.
So don't assume it's all nerds having fun. Now, is it what the future will look like for everyone? I don't think so, but that's because I expect AI is bringing us a more unified dev experience, with more developers than even Google has. This will make general programming harnesses very good, quite fast, solely because said harnesses can be differentiators in the AI race. So we are seeing some of the largest companies inthe world, with the largest research budgets, dedicating more money to the dev experience of their product than almost anyone else does.
You're talking about the best businesses. Arguably the article is talking about the mean of the "startup" tranche of businesses.
Google isn't a SaaS. That's (is or is tangential to) the kind of businesses this article is referring to. And they make decent money, they just aren't a unicorn that didn't die.
(This is temporary until the AI gets better judgement than humans, then capital will therever be the most powerful moat in a market full of dystopic, consequentialist, incredibly long-sightedly-greedy companies)
Long before A.I., tiny 10 person startups have been able to revolutionize industries.
Why wouldn’t capital simply lose its value even further, especially as capital globalizes further
The article argues (and I think I agree) that how you choose to incorporate agents into your work will be a differentiator and most great companies will have a unique take on it.
That's true for any product.
Who are they going to sell the product that this whole harness would produce? Any ideas?
1. The quality of Mattermost and Zulip was NOT exactly as polished as Slack.
2. Deployment of such stacks was a problem. A complicate operational overhead.
These days, that's no more the case. Agents can:
1. Make pixel perfect clone of Slack or Jira with SQLite or ejabberd behind.
2. Or they can deploy the Zulip or Mattermost or GitLab for you just give them an SSH key to the machine and see them bringing the stack to life.
So now and back then are not the same.
If it's that easy, then maybe you can do it and give the world a free and better-than-Zulip alternative to Slack?
Software in individual service doesn't have to scale, secure or maintainable.
Software is a thruway artefact now.
People are having hard time realizing that.
I have to admit, I really am having a hard time realizing that.
Mattermost isn't actually all that bad and its mental model is pretty close to Slack, at least compared to Zulip. I actually prefer Mattermost to Teams, and would view it in the same ballpark as Slack.
> Deployment of such stacks was a problem. A complicate operational overhead.
Not at all! Their stack is actually very reasonable and quite easy to setup: https://github.com/mattermost/docker/blob/main/docker-compos...
Nowhere near the nightmare that self-hosting Sentry is like: https://github.com/getsentry/self-hosted/blob/master/docker-...
I'd say that in Mattermost's case the group calling functionality was locked behind a subscription as well as some other stuff last I checked, which makes it dead on arrival for many. It was quite nice software though, even before being able to vibe code your own (though tackling videos sharing and RTMP will be anything but trivial).
> Make pixel perfect clone of Slack or Jira with SQLite or ejabberd behind.
There's A LOT of functionality and features in Jira, the only thing that saves claims like that is that you probably don't use 80% and can just build what your company needs. I more or less did that for the hell of it (MariaDB + Garage + Dropwizard/Java + Angular/Spartan) and there was still so much supervision an changes that were needed - after pulling the Jira DB scheme and throwing about 60% of it in the trash (didn't need the automation), there was still A LOT of stuff to do and even the plans eat up a whole bunch of the context of the coding tools you intend to use, no matter how many checks and scripts you write to automate guardrails for when the models simply don't recall everything that is relevant.
For people who want to self-host something basic, there's already the excellent https://kanboard.org/ (very very fast)
For those that want something similarly depressingly slow to Jira while still locking some stuff behind a subscription (at least in the versions I tried) there's the passable https://www.openproject.org/
If those don't fit and you don't see anything exactly like what you need, you might try to build your own. But yeah, not a weekend project.
I think you missed the part where Google was successful because it had the best engineers, and then went all Alphabet.
There is such a thing as too big too fail.
When the definition is that broad, aren’t you effectively already defining what SaaS is today.
Especially when you include interfaces in that definition.
I feel this is probably coming and perhaps inevitable, but I fear for the poor employees this sort of thing will be tried out on first.
Also https://web.archive.org/web/20260725174355/https://marshallb...
The harness controls the AI input and output into an outcome that requires judgement.
The human part is that judgement. Not the decisions - we can pass that off but what is a 'good' thing?
Tomatoes in your fruit salad? Tomatoes are fruit. Olive oil in your engine? Lubricants are lubricants.
AI slop creates software and documents and websites very quickly. It passes the tests. It does the thing but can it be trusted? Consistent positive judgement calls create a track record and that creates trust.
This explains the anticipation about Jev.
Trust is now what sells to the highest bidder.
Because someone was so sick of expensive Adobe Photoshop they created their own vibe coded PhotoShop for $2000 worth of tokens and then selling it on for way cheaper. [0]
Some YouTuber got so sick of Adobe Premier they literally vibe coded their own full fledge video editor exactly to their needs with everything else that they don't need stripped out. [1]
I know of a case where a totally non technical person sitting in an south asian city wrote his own financial management software for his business in just three weeks. Vibe coded and now his daily driver tracking accounts, payables, receivables, contracts, parses PDFs from his inbox, populates forms and full workflow that HE needs for his specific commission/resale/distribution business.
So what this company is going to sell with that Harness? Or would it be selling just that Harness to other companies like those Rails bootstrap SaaS boilerplate businesses that used to charge $300 for that junk with lifetime update promise?
EDIT: Formatting
[0]. https://www.reddit.com/r/vibecoding/comments/1wpaies/my_vibe...
You can never prove that there aren't any bugs. The definition can also change according to time and expectations.
I've made tens of internal developer tools because I couldn't gaf about using some third party vendor, all before vibe coding. Even then we purchased software from vendors, and we still do it now.
You can't cherry pick 2-3 pieces of problematic software, imply that it's how everything works. You're vibe coded Photoshop will never be used seriously, unfortunately. Same goes for the video editor, if you're fine with an autonomous AI agents listening to some internal github repo for issues that get auto or human reported and attempting to provide live fixes and feature enrichment, you aren't going to get that.
The problem really does stem from the fact that these agents, no matter how good you keep making them are still limited by a time horizon. AI companies have gotten really good at extending that horizon. The contexts get longer, prompts get less specific but theres a deadline to all of this. Take that LLM point it to a specific problem let it run, pull it out start new. The LLM will inevitably accumulate technical debt in it's own context, that starts to create a rot. You can try to improve this problem by making sub agents, etc. but you will 100% run into context rot after a period of time for the lead agent.
New technologies will always lower the bar and raise the ceiling.
We may not be able to see the ceiling yet while still learning capabilities and implementations 5-20 years from now.
Also, I can't help but laugh a little that for the most part, text files are being called a harness.
Their main thing, iirc, was detecting ~abnormal~ looking emails. It's the main way companies lose money and IP, it's not via crazy hacks, it's obviously the human element that is the biggest vulnerability in all companies. A stressed employee gets an email saying hey i'm the ceo i'm in a rush, or hey it's totally jennifer from your team i'm stuck outside, send me all the files/do x quickly. I don't have any fucking idea what they're doing now and I bet you they dont either
This must be a response to the cyber stuff from the big labs
Software factories don't exist is an easily falsifiable statement https://github.com/topics/software-factory