Where’s My Saddle?

Photo by Roger van de Kimmenade on Unsplash

Today I had a really great chat with a coworker and friend I’ll call John, just to protect his identity. John and I met, I want to say, more than seven years ago — I believe the first time we actually talked was when he was interviewing for a position on one of my teams. I was very impressed by his work ethic, his intelligence, and the fact that he seemed like one of those serious, analytical people. He struck me as a software engineer who wasn’t just good at what he did, but genuinely steeped in the technology — someone who could talk your ear off about why certain programming languages were better than others, or different coding paradigms, or why one code structure beat another.

So I hired John, and he joined the team. I remember we actually met face to face once — and I should probably pause and explain why that’s even worth mentioning, since it might sound odd that I’m noting I actually met someone I hired.

The company I work for hires people from all over the world and you don’t need to be in the same room or office as everyone. I’ve managed as many as 115 people, at one point as a director — through managers and senior managers who reported directly to me — all from the comfort of my home office. We hire people to work remotely, and yes, after COVID there was a bit of a kerfuffle about whether remote work was hurting company productivity or the quality of people’s output — which is kind of ironic, given that this company has been around long before COVID, built from the start on contributors working from anywhere, any time of day. But I digress.

So I got on a video call with John, and the thing that really struck me today was where the conversation went. We started on work, but ended up talking about something we’re both noticing — not just at the company, but in the industry as a whole.

Imagine a domain where everyone who’s worked in it long enough has settled on a few things as simply true — the way everyone knows water boils and freezes at the same temperatures, give or take some factors that could nudge those numbers slightly, but nothing that changes the basic rule. Axioms. Things you just take at face value.

In our industry, one of those axioms is that throwing more people at a problem doesn’t actually solve it faster. Adding more engineers — or “resources,” as I hate to call people — to a software problem doesn’t necessarily speed things up. It can actually slow things down and cause real frustration. Anyone who’s been around this field for a while knows this as fact.

And here’s what John was worked up about today: we’re seeing more and more people believe that, with the rise of AI and everyone having access to agentic tools, you can just throw anyone at a problem, because AI will lower the barrier to entry and get people productive on day one.

It’s tempting to believe, because the latest LLMs really are capable of impressive things, and they can genuinely take someone from zero to some level of competence without the usual ramp-up time — especially for someone seeing a codebase or project for the first time.

But reality still shows, even with today’s most advanced models, that it takes a while for anyone to actually feel productive using AI to solve the problems leadership thinks can now be solved instantly. Everyone I’ve talked to who uses AI daily for typical software engineering work says the same thing: they’ve stopped writing code and now spend most of their time reviewing what AI produces. They’ve become reviewers for AI — catching the mistakes and slip-ups it introduces before code gets submitted.

And because these tools are so fast and can juggle so many tasks at once, the sheer volume of code being generated is so high that even the best-intentioned humans can’t keep up reviewing it and live up to the expectations from leadership, assuming that because AI codes fast, the human in the middle should be able to keep pace and ship just as quickly.

But can you imagine reviewing a few hundred, or a few thousand, lines of changed code and making real sense of it in a couple of hours? I don’t think the human brain can hold that much context — not unless you spend days, weeks, maybe months tracing every connection and every way a given change could ripple through the codebase.

So everyone I’ve talked to saying they’re spending their time reviewing AI-generated code also told me that it’s slowly becoming acceptable to just trust the AI and do a perfunctory pass instead — especially since AI is also writing the tests meant to validate its own changes. By the end of the day your eyes are blurring, and you do your best to review what you can, hoping against hope that the AI-generated test harnesses catch anything major that slips past you.

So we’re back to throwing engineers at problems, hoping to ship faster to customers who supposedly can’t wait to get their hands on the latest version. Except the reality is that customers don’t actually like to upgrade that often, and they’re often a little reluctant to try something brand new — preferring to wait a few iterations and let other people shake out the bugs first, so that by the time they install it, the software is a bit more stable and reliable. It’s a myth to think paying customers are eager to install your product every day, or even every week — unless you’re offering it as a service where the user never installs anything at all. Nobody’s clamoring to upgrade their systems every time you, or your AI, push out new changes.

John was genuinely worked up talking about all this. And I should mention — he’s normally one of the most polite, careful, philosophical people you’ll ever meet. Picture the quietest person you know, someone who expresses complex ideas fluently and calmly. That’s John. But today, maybe for the first time since I’ve known him, he was dropping F-bombs left and right out of pure frustration. “These people are insanely crazy,” he said multiple times. “Absolutely f-ing crazy!”

It was almost funny to me, because I like John a lot and I’d genuinely never heard him cuss before — but it also showed just how irate he was that here we are, in 2026, and people who’ve done this work for decades still fail to see the irony of repeating mistakes they’ve known better than to make for years, just because AI looks like the silver bullet this time. Despite how good these models are, it’s genuinely stunning how quickly people throw caution to the wind, hoping AI can somehow override what decades of hard-won experience have already proven true.

Before we hung up, John shared something that made me laugh. It was a proverb he couldn’t remember the exact origin or wording of, but it went something like this: if one person calls you a horse, you tell them they’re crazy. If two people call you a horse, you figure those two people are crazy. But if five people call you a horse, it’s probably time to go buy a saddle.

Written between meetings, edited between cups of coffee.