In his post Why Software Factories fail the CEO of HumanLayer argues that a combination of reasons (agents, training, harnesses) are the reason that we don't see effective "software factories".
But I disagree with the whole premise: that a software factory is possible or desirable.
The reason software factories fail is because software cannot be made in a factory. You're not making widgets that conform to a spec, you're building an experience for a human. While his company is named well enough, I think his post misses the point that better training can never tell you if the thing you've built works or not.
You—the human—must ride the ride. Click the button. Feel the wind in your hair.
Because as we all know, language-to-code specification is inherently lossy. You can't pack all the necessary details into code without... y'know... coding.
So while it's great that we're building cool things faster than ever, you can't think of agentic software engineering as a "factory" or else you'll believe the wrong things about its capabilities and end up driving too close to the edge. You avoid shipping slop by believing the right things about who software is for and how it's built.
As a brilliant co-worker of mine once said:
Let robots do the stuff robots are good at.
Robots aren't good at crafting a human experience.
Only you can do that. ❤️