Confession time. My last post, Reviving This Blog, was written by an AI. It has a “Full disclosure” paragraph explaining that an agent did most of the migration while I made decisions as it grilled me.
I’ve never been good at writing. Essays were my worst subject in high school and college. I liked science, math, doing things with my hands. However, any type of structured writing, outlining, organization, has been met with the mother of all dopamine crashes. So, when I refreshed the site this summer in my first attempt at trying to get better with agentic tools, I handed the words to the robot too. It wrote something that looks fine…buuuuut, AI writes weird. It’s rooted in patterns, naturally, and once you see them you can’t unsee them.
Spot the robot
Go back and read it. Or don’t, I’m not your manager. I just wanted to specifically call out that the post opens the second sentence with “Then the world got interesting,” which is how a language model says COVID without saying COVID due to its content filter. It also chose some other interesting phrases, like the blog had “lost every prioritization battle it entered.” And Netlify “has earned its keep.” Nobody I know talks like that, let alone about a static site host.
Then there’s “I’ll take the hit on that one.” That one stings, because it’s actually my phrase. I keep a voice guide for these agents so they write like me, and it says I own misses with “I’ll take the hit.” So the agent read the guide and dropped my catchphrase right on cue, like an actor nailing their line. Except, it’s my line in somebody else’s mouth.
My favorite is the closer. “I’m not promising a cadence. I’m promising the next gap won’t be six years.” Not this, but that. It might be the most AI sentence ever constructed. It sounds wise and also it fits in a tweet. To be fair, it was right. The gap was four months, so the robot set the bar low enough that even I could clear it again.
It gets better though. Before that post shipped, it went through a second “voice pass,” where another agent session edited the first one to sound more human. Robots reviewing robots, to sound more human. Gives pure :hmmyes: (🤔) emoji vibes. Topped off with me clicking approve on the PR without actually fully reading the full blog or new site code.
And the whole reason that post existed was that my bio had drifted out of date and the site couldn’t agree on where I worked. As of today, the about page still says I’m an engineering manager at Lambda. (I’ll fix that soon after this post) I work at SF Compute now. So we’re right back where we started, and I’ll take the hit on that one. (See? It’s catchy.)
So what changed
This post started as two voice memos. I recorded the first one at home on October 26th, and the second on November 7th while driving to Detroit Metro for KubeCon. I ramble into my phone, Apple’s Voice Memos transcribes it with decent accuracy, and a model cleans it up lightly, cutting the filler and the “um”s (there are a lot of “um”s) and grouping related thoughts while keeping my phrasing. The rule is that at least 80% of the words have to be mine. Then I build the post from that. The thoughts come first, organically, with the human in the loop before AI is introduced, instead of the other way around.
The irony isn’t lost on me that I’m using AI to help write a post making fun of a post AI wrote. That’s kind of the point though. The tool isn’t the problem. Where the human sits is the problem. We cannot lose our critical thinking or creativity in the process for the dopamine hit of “producing more.” More slop that is.
Off, on, and in the loop
Kief Morris has a good article on martinfowler.com, Humans and Agents in Software Engineering Loops, that splits software work into two loops. There’s a why loop, what we’re building, and a how loop, how we’re building it. (In my voice memo I called it the “what loop” and admitted I couldn’t remember which one was which. Organic human thought, warts and all.)
The why loop is still a human loop. You work with a product manager, you’re doing user stories and gathering requirements, sometimes working with a TPM to get good definitions of done and break down the work and scope. The how loop, what language and what technical implementation, has usually gone to junior and mid-level engineers, while senior architects work the why with PMs and TPMs. Vibe coding has really taken over the how loop. Now someone can give a prompt to an AI and say “I want you to build this,” and the AI will go and loop and figure out on its own how to build it.
That leaves three places a human can sit. Humans off the loop is vibe coding, where you let the agent go off on its own and loop endlessly until it finishes. Humans on the loop means you’re aware and guiding, either through skills or by giving it proper context and guardrails. Humans in the loop means you’re interacting with the agent the whole way, and it stops and checks in. Much like Claude before Auto mode.
My last blog post was humans off the loop. I gave the agent a prompt and a voice guide and approved what came back. Pure Vibes. That’s fine for converting old reStructuredText posts to Markdown. It’s not fine for the part where I’m supposed to be the one talking and giving my thoughts a place to exist in the digital world.
The software factory
The same thing is playing out at a much bigger scale. There’s a lot of noise on social media right now from software engineers sharing how they’re harnessing AI. Not best practices, just practices… (get it? Just vibes…), with everyone trying out different things, and honestly it’s a lot of fun to watch but also the cognitive load is too much at times. One term that keeps coming out of it is the software factory. That’s multiple agents tailored to specific tasks, handing work off to each other in an automated, assembly-line style, sometimes running entire SDLC (Software development lifecycle) pipelines or even whole companies.
I’m skeptical. Maybe that’s my background, but I’ve seen how off the rails AI can get unless you have a well-designed harness. Even then, nobody knows what a well-designed harness looks like yet (though some of the posts from Lauren Tan are intriguing), because a harness that works for one model might constrain the next one with how quickly these models are advancing.
One of my coworkers and good friends, Billie Cleek, said the software factory reminds him of the no-code era back in the early 2010s, and I’d agree. Agents can run 24/7 and hand work off overnight while you sleep, but that doesn’t always mean they should. From what I’ve seen, even from the people building these things in public, there’s a lot of time and cost and tokens spent to get to the same result you’d get with a human more involved, especially when you’re trying to enshrine good software engineering fundamentals. Lauren Tan (@poteto) at Cursor, whose setup some people call a software factory, has said she doesn’t align heavily with the term. It’s good fundamental software engineering: design, user stories, requirements, development, validation.
Don’t cut the juniors
This is the part I actually care about. A lot. As a former line manager for an internal platform team, I’ve seen people come to believe this boost in productivity fundamentally changes how we design engineering organizations, how many people you need and what those teams look like. I think we’re seeing an overcorrection toward very senior engineers, folks with strong fundamentals who have worked somewhere with good structure and seen what good looks like. Some folks believe AI is replacing junior engineers. I believe that’s fundamentally flawed. I’m also hopeful this is just coming from influencers and non-tech VC founders who are chasing the current craze hoping to make entire profit margins with no OpEx (Operational Expenditures).
Those senior engineers are going to age out. Not immediately, but they’ll retire, and you’ll have no one to promote into their place. Junior engineers need to be taught how to be skeptical, how to review, and how to understand system design, the same way we were taught to verify how code is processed in the underlying systems, in the kernel, built with compilers. Do you need to understand exactly what machine code the compiler is outputting? No. But you need an understanding of the layers of the stack.
We’ve been here before. When the first compilers and higher-level languages came about, people had the same argument we’re having now. You can’t trust the compiler, you have to verify everything that comes out. Eventually the compilers got so good that we learned to trust them, and we learned to test and verify the code instead. The same is going to happen with AI generated code and hopefully is already happening in some orgs.
We didn’t push out junior engineers back then, so why are we doing it now? I don’t disagree that AI most likely changes how many people you need to do the same amount of work. I don’t believe it changes the composition of the team. You still need young folks who understand this new world we’re moving toward, and we need to be mentoring them on how AI hallucinates and where it falls short, so they grow into the humans on the loop.
Signing off
That’s two voice memos’ worth thus far. I like where this is going. I’ve felt more organized since I can remember. This feels natural given the raw struggle to organize my frantic mind in the past, before AI. I wanted to do these daily and the habit hasn’t really stuck, which is par for the course for me.
And if my next post opens with “Then the world got interesting,” you have my permission to call it out.