July 23, 2026
Ctrl+Alt+Delusion
Why Software Factories Fail (or: harness engineering is not enough)
AI was supposed to build apps alone — commenters say it’s mostly making bigger messes
TLDR: The article says companies are rushing to let AI build software with less human checking, but that extra automation can’t solve deeper quality problems. Commenters split between cheering the reality check, mocking the hype, and roasting everyone involved for feeding the same “slop cannon” they now complain about.
The big promise of the moment is deliciously simple: let artificial intelligence write the software, skip the boring human checking, and watch products fly out the door. But this post throws a bucket of cold water on that dream, arguing that you can’t fix shaky AI coding by just wrapping it in more rules, more tests, and more automated loops. In plain English: if the brain behind the bot is flawed, no amount of fancy supervision will magically turn it into a flawless worker.
And the comments absolutely went to war. Some readers praised it as one of the clearest explanations yet of why these systems keep tripping over themselves, with one saying online debate would improve overnight if more people understood how these models are actually trained. Others were less charitable and accused the author of throwing stones from inside the same glass house, with the brutal dunk: “you aren’t part of the slop cannon, you are the slop cannon.” Ouch.
There was also a mini side-quest in the replies, where one commenter smugly noted that the missing details about StrongDM’s “lights-off” coding experiment were actually easy to find with an “old-fashioned google search too, no deep research agent needed.” That little jab pretty much summed up the mood: half serious concern, half meme-fueled eye-roll.
The biggest shared anxiety? Reviews are getting worse, buggy code is getting through, and teams may be moving faster only in the same way a shopping cart speeds up downhill — thrilling right until it hits a wall.
Key Points
- •The article argues that current enthusiasm for AI-driven software factories is centered on increasing automation and reducing human involvement in coding and review.
- •It cites StrongDM’s 'lights-off software factory' and OpenAI’s Symphony as examples of organizations promoting software-factory approaches.
- •The article references warnings about outages and code quality decline linked to coding-agent use, including remarks from Mario, Matt Pocock, and a Faros AI report.
- •The Faros AI report is described as showing lower pull-request review quality, more unreviewed merges, and increases in incidents and bugs per developer, though the author notes it is correlational rather than conclusive.
- •The author’s central claim is that additional harness engineering and prompting cannot fully solve software quality issues that stem from how coding models are trained and evaluated.