I can't help but cringe at the "cost of code is now zero" meme that this article repeats because in my experience the biggest cost of code was always the activities around the code - planning, communicating, reviewing, validating. This author, despite repeating the meme, seems to agree. Their emphasis on writing PR descriptions and specs for humans rhymes with my experience and it sounds like a nicer way to work than chasing some dark factory fever dream.
I also thought the "Two Modes of Working" section was useful. People get wildly different results from coding agents depending on how they use them, but I've not seen a lot of actual guidance on when to do X vs Y.
Personally, I've been using what the author calls "sidekick mode" since last October - before the supposed "AI got good now" threshold - and agree it's a more useful default for an engineer than "delegate mode".
Writing code is the easy part. Figuring out what your new feature or product should do is the hard part.
Maybe this is a paid feature that we're going to charge money for. Or we're going after a new market segment, or trying to take business from a rival. A new feature might have regulatory or legal requirements, somebody needs to understand all that.
At the very least, someone wants this new thing for a reason. You can't just guess, you actually have to talk to people and understand the problem domain lol.
I almost always have a pretty good idea of what I'm making and how I'm going to make it before I start.
I’m starting to come around to the belief that vibe coding isn’t engineering. Not saying Ashby is vibe coding. The post certainly says they’re super serious about making sure everyone understands every line of change. I wonder how they manage that in practice.
There are few empirical studies on the effects of informal code review, what most of us do on GitHub and Codeberg, on error rates. It’s not much. But a little side line mentioned that the effect we do have on error rates disappears the more code you read per hour.
So you used to have the 80/20 rule. 20% of your dev team is doing 80% of the work. Now we have a tool that enables the 80% of the team that manages to squeak by without doing a whole lot to… write a whole ton of code that the 20% now have to review… only if they read too much it’s not going to do a whole lot.
So I dunno. Seems like that combined with the come wave of price hikes we might see orgs talking about rolling back their AI usage later this year.
We only hire engineering managers that are deeply technical. This gives us more ability to work closely with engineers and make sure they're using new tools in the ways we want. It's not perfect. Sometimes, it goes wrong. But, that's why we continuously add tests and other tools to help us spot mistakes.
The 'blip' post-AI is up ~30% for the months in question. There simply isn't enough data here to prove or disprove the thesis that "customer issues remain broadly stable" - you could equally argue something more along the lines of "AI engineering does not increase issues under ideal conditions but amplifies issues under external pressure."
One other point of note is that the chart only goes back to January 2025, and only dilineates "majority AI" from "not majority". If the usage rate in January 2025 was at 45%, "issue rate remained roughly stable (with a giant asterisk) when moving from 45% to 51%" is not exactly a compelling story. There's no narrative I trust less than one created by cherrypicking data to lend itself an aura of faux objectivity.
I think you're right that the data presented doesn't prove or disprove the thesis that "customer issues remain broadly stable". I think it lends credibility to the idea that "AI didn't lead to an obvious and immediate slopfest of bugs".
As for the bumps, we've noticed them annually and we haven't yet found a solid reason why. There’s no correlation with customer growth, PRs shipped, or features shipped. It's likely a combination of various seasonal effects in our own hiring and customer onboarding. What isn't in this article: we've more than doubled our customer base in this time period, our customers are more mature and raise bugs more readily, we've got ~60% more engineers, and our product surface area has grown a lot too. I'm yet to land on a satisfying metric that normalizes bug rates in a changing environment.
The cost is heading to 0???? Do they not pay for Claude, Anthropic, OpenAI? Have they not seen the sudden price increases? Hosted locally LLMs still have electrical and server costs.
I believe not. It shows they dont truly understand the nuance of what they are talking about.
I have no reason to doubt the sincerity of the authors of that post (and enjoyed reading it), I just hope they have a good sense of how they're recognizing responsible and irresponsible use and taking that into account in a perf process.
E.g. in a recent review I wrote I picked up on an engineer not splitting up React components aggressively enough and over-relying on deeply nested ternaries. In that same review I also commended the engineer for being particularly excellent at explaining complex bugs to our support people and called out one example of a nasty race condition.
Wait, whoever is reviewing the engineer isn't already familiar through exposure to – if not participation in – flow of work?
But, like, this is a model you can embrace without AI. Have some people make messes, have others clean them up.
To answer your question: I mainly used Lever before and it was more or less OK, just like Ashby is OK. So, I'm not saying it's a bad product, it's probably a good product. I just don't see it "so much better" to justify a switch from Lever from my point of view. Both are "good enough" from an hiring manager perspective (or at least, my perspective).
This (correct) thesis should illicit an interesting question about the future of SaaS markets: What happens to the SaaS markets when the cost of code approaches zero?
Coincidentally, the company that authored this post is a perfect case study.
Few people truly grasp how hard bringing a software product to market is. The feature density required to even begin selling your product often requires more human capital than early stage investors are willing to underwrite. Some founders have the background to command those terms, but most have to wedge into a market with some very small insight.
I was at a dinner with the CEO of Greenhouse in 2021 and vividly recall him explaining just how deep the feature set of a new ATS needs to be to even enter the consideration set of a buyer.
One serendipitous way companies could enter these markets is by being founded in one market, and pivoting into another. Ashby didn't start as an ATS. They were originally targeting a mid-market ERP (a hot thesis at the time). Ashby is a prime example of a company that likely could not have entered their current market because of the sheer engineering resources required to break through (and kudos to them for doing it!)
But as the cost of code approaches zero, the deterrents of entering the market also drop to zero. The next Ashby could just pursue ATS from day one. In a sense, this is the entire story of humanity and especially software (consider the cost of building an ATS in 2012 when Lever was founded vs the cost Ashby faced in 2019).
But the slope of this change is unlike anything before it.
ATS, like many other markets, has a handful of big players and a long tail of smaller offerings. No matter how many long tail providers there are, the economics of entering a market suppressed that number compared to what can and will exist tomorrow. The cost of building an ATS a decade ago was literally orders of magnitude more than it is today.
In 2016 I heard a A16Z partner describe a future where every software market would have offerings in virtually every little niche one could imagine. Years before the LLM renaissance, this sounded insane. How could the market afford to build and sustain each offering?
I don't think companies are going to start building their ATS in house, but I do believe that cost of producing code is approaching to zero, and that means hundreds or even thousands of offerings will exist in every shape, way and form we can imagine.
Are these related? The difficulty of bringing a software product to market has never been that it takes time to produce code. We’ve had the concept of MVPs for decades and pre-AI companies have raised hundreds of millions off of no code prototypes.
YC has always preached that you launch. Launch today. Launch as soon as is humanly possible. And learn. And iterate. You get something into the market as soon as possible. Spending years building the perfect software has never been the strategy pursued by technology startups, and not because of the time it takes, but because you need users to know what to build.
I think we could even go as far as to counter your argument. We know that constraints are powerful. Constraints force us to focus. Imagine you decide to build an ATS. You spend a bunch of money on Claude Code to generate hundreds of features, every little idea you can think of, you include it. You launch with the most complete ATS on the market. Ashby can’t hold a candle to the depth of your software. Do you have a better chance of success than if you had launched with a much smaller surface area, experimented, found the right area of focus, and iterated according to user behavior?
I encounter a lot of vibe coded software products that people are launching and I am not precious about code, a good product is a good product whether it was hand crafted over a decade by a thousand well paid programmers or generated in an afternoon by Claude Code from a laypersons prompt. I can name one single vibe coded product I pay for that is genuinely good software. All that the AI age has done is show that most of us absolutely suck at building products, and that even with free code, what we produce is garbage.
“How could the market afford to build and sustain each offering?”
I think you’re conflating build and sustain here. Nothing fundamental about business has changed. Sustaining a business isn’t about being able to generate code. Even in a world where code is free, how could the market sustain 10,000 Ashbys? Where are the customers coming from?
As a concrete example: how many customers does Jira have despite Linear being better software in every single way? Would generating 1,000 more Linears change Jira’s market position? Or Microsoft Teams, Teams is awful, there are so many better options. Will generating 1,000 more better options change Microsoft Teams’s position?
This does not mean that R is on the only cost (T-R grows over the life of the company). It does not mean that, even if R were 0, you could launch a product. But R is a real cost.
What proportion of T does R need to be to see thousands of Ashbys appear?
Ashby is a good example to discuss because it is serious software for serious businesses, it isn’t a calorie counting app. Serious business expects so much more out of their suppliers than software, Ashby is so much more than a bunch of code.
So, a thousand people generate their own ashbys, then what? Are companies like Shopify and Ramp going to trust their ATS to some person who generated some software in a day?
I agree that the cost of building software has gone down a lot and that lowers the barrier to entry for building a software business. I completely disagree that the market could support even 10x the number of companies in each vertical, let alone 100x or 1000x.
You could clone Ashby today and go to Ramp and offer them a 50% discount on whatever they’re paying Ashby and there is zero chance you win their business. You could clone Ashby and the add 10x the features and go to ramp and offer a 75% discount and there is zero chance of you winning their business.
the reality actually is, you DO need depth in order to close lucrative enterprise contracts.
but startups will use anything. you use early money/traction to fund the deeper features. that's saas/startup 101.
anyhow, saas just means the second "s" matters more. maybe software gets cheaper (that's actually an unproven hypothesis), but service encompasses many dimensions. the most lucrative contracts won't be eaten by fly-by-the-night operations almost by definition. any startup that knows the words 'vendor risk' will tell you.
It’s a product that feels like it’s designed and built by the C or D team. The slowness and bloat of Rippling with wild, fever dream interaction models and workflows that frustrate and confound.
(And enjoy fucking emailing them if you need anything at all: almost literally no self service paths exist in the app for account upgrades, adding on SAML etc. And I do mean that you have to email them—last I checked it’s a mailto: link, not even a form.)
We had to build a lot of product in a short period of time to be a viable option to replace incumbent ATSes who had fought off many startups (and it couldn't be buggy or unreliable). The startups that focused on simplicity and usability just never made it. So, we focused on building something flexible and customizable first, with usability being something we wanted to make excellent second. You probably felt that over the past few years, and I'm sorry you did. We've been working hard to improve it, especially in the past year (e.g., you can self-serve account downgrades and some product upgrades now!).
Still more work to do, but thanks for putting up with some of our deficiencies :)
"We must think more" - unless your industry wakes the hell up and standardizes its work. In that case you don't have to think more, because you have reliable standards that remove the need to think. I hope I'm not dead before the industry finally grows up.
UPD: surprisingly not