The advantage that software always had was always that it scaled well.

A team of ten engineers can build a thing that works for a team of ten experts in a specific field, but, nobody gets rich doing that. The same team of ten engineers can build a thing that works for ten million consumers around the world, and while your staff costs haven't changed, your potential addressable market has, by several orders of magnitude.

A significant challenge to working at scale (there are many others) is that of quality. A team of ten experts using a custom built tool may be able to tolerate that the software fails one in a thousand times. But, if you have ten million customers, that would mean ten thousand customers write you an angry support ticket every time they use your product.

So, quality is one of the vital properties that needs to be built into software that's intended to scale.

Agentic software development (or, if you want vibe-coding) has, to a certain extent, flipped the scale/engineer proposition around.

You no longer need a team of ten engineers to build a thing that a team of ten experts use, in fact, they can quite possibly vibe it up themselves. They'll probably be quite happy too, maybe it fails one in a thousand times, but when it works, it does exactly what they need and they don't need to try to make a team of software engineers understand their precise requirements.

But, a vibe-coded application launched to millions of users will almost certainly be found to have multiple failure modes that weren't found in testing, probably within minutes of launching (waves to Epic games).

I expect, this will lead to a different way in which software as an industry is both hollowed out, but also democratised. Small businesses with niche needs will no longer hire developers to write their internal tooling. (I suspect a small industry of consultants fixing specialist tooling will thrive however.) Large, at scale, organisations will decide it's cheaper to code things by hand though, quality problems scale up with your user base, and agents are still not yet that good at creating complex software with the kind of quality you need at large scale.

So, there will be fewer small-scale jobs, fewer entry-level jobs, and the jobs that remain will be the hardest – but also the most valuable. I find myself with a new reason to be happy I'm not trying to join the tech industry as a junior developer. If I was looking at training as a software engineer though, I'd be looking long and hard at my education choices. Anything that doesn't cover the fundamentals, doesn't deal with the challenges of building for scale, and doesn't emphasise quality, probably won't give me what my potential employers will need.

So, good news for experts in niches with software needs, bad news for software engineers looking for niches.

There are, of course, a bunch of reasons to dislike the current revolution. The disruption to a job I enjoyed doing, the source of that disruption and their motivations, the ever-so-slightly sub-par quality of the work AI does. There are reasons outside of tech to really dislike AI, but none of those have been quite what's been bothering me.

It's the chaos, the “everyone's a coder now” liberty that means now everyone has not-invented-here syndrome, and there's no practical barrier to stop you inventing whatever “it” is, here, again. From writing your special, one-weird-trick AI Skill.md that makes you so much more productive (citation needed) or so much more cost-effective (citation needed), to creating your own tool that you always needed but couldn't get anyone to sign off on. The overwhelming, disorienting, what-is-it-this-week, chaos. Everyone's got their own thing now, nobody bothers to work out how other people are solving this problem because they can just solve it themselves with AI, and that's quicker than spending half an hour doing the research.

Which leads me to the primary use-case I can see people using AI for regularly. The thing that everyone seems to be doing all of the time.

AI lets you circumvent your employer's vendor approval process

With, it has to be said, as side order of letting you avoid doing all that nasty research into whatever the problem is you're actually trying to solve and whether it's already been solved.

So, now, instead of deciding to build or buy it's quicker, easier, and so much more satisfying in the short term to just “build”. AI will give you a personalised, custom version of whatever it is you need, inferred from a bunch of times other people have already solved this problem, without ever having to deal with researching either the problem or the solutions. You don't need to get the vendor approved, because your employer has already signed off on your AI subscription, and most likely your cloud hosting provider. You can whip up something exactly suited to your needs in an hour or so and be on your merry way.

Except,

So can the person next to you,

And the person in the office two doors down,

and all of you can do this all day long, everyday.

So, congrats, you have a beautiful stock management tool that you're using to track your macguffins, and now your employer is running a dozen custom-made tools for all sorts of use-cases and everyone's so much more productive. Your finance department are thrilled that they can stop denying purchase order requests, because that was a hassle, and, let's face it, they were a week away from writing their own tool to review and automatically triage all requests for new tools or vendors needing approval.

Smash cut forwards a few months and you're a bit peeved that nobody is tracking their macguffins in your tool, but the folks two doors down have written an MCP for their tool so now everyone is just asking ClodGPT to track the macguffins and you only discovered that you're not the only one when somebody asked you for an MCP so they can add your tool into the context of their new agentic workflow.

Smash cut forwards a year, and half your companies (rapidly growning) token budget is consumed fixing bugs, shipping patches, and injecting extremely niche functionality in the dozens (and multiplying) internal tools it now maintains, each with their own very specific slice of the overall picture. Probably all talking to each other via exposing an MCP and having “agentic workflows” ferrying the data between systems on a regular schedule. Fortunately for everyone involved, none of the tooling works so well that anyone lost their job, but, you do spend all day asking computers what they're actually doing.

And, I'll be honest, I hate that for you.

But, you'll probably think it's great. You've used AI to solve all your problems, create some new problems and then solve them again, and you felt so productive doing it.

And, that, that last part, that nagging suspicion that you'll be so much happier like this.

That's what really makes me sad.