Posts

For most of the history of enterprise software, customization has been a sign that something went wrong. The ideal product was the same product for everyone. A vendor built a broad application, added configuration screens, exposed an API, maintained a plugin system, and gradually accumulated enough features to satisfy a market. Customers adjusted their processes to fit the software, or paid consultants to close the remaining gaps. That model made sense because software was expensive to write and even more expensive to maintain. ...

I’ve noticed that many developers (myself included) experience a certain anxiety when using AI coding agents. The model writes tens of thousands of lines, now with a coherent architecture, consistent naming, decent design, and increasingly resembles what you’d expect from a mature codebase. It uses practices you weren’t always aware of. The natural reaction is to compare it to yourself and think: “I could never have written this.” Perhaps that’s the wrong comparison. ...

The following is a narration of our own experience implementing a software dark factory: a software system that autonomously implements software. We build the factory; the factory builds the product. This is our own experience applying the principles outlined in Ryan Lopopolo’s OpenAI post on harness engineering, which may well go down as one of the most influential software engineering posts in recent times. There is also a Latent Space episode with @swyx interviewing Ryan Lopopolo, which is well worth listening to alongside the blog. The core framing from the post is simple and powerful: ...

How to Build Software in the Age Where AI Writes Most of It There’s a lot of noise right now about agents building software — autonomous systems, AI engineers, fully automated development pipelines. Some of it is true. But most of the conversation misses something fundamental about how software organizations actually work. If you want to build software in the age of code generation, you shouldn’t start by thinking about agents. You should start by thinking about the structures that have survived for decades, and asking why they lasted. ...

To me, the real inflection point came at the end of May 2025. That’s when Anthropic—the company founded by ex-OpenAI researchers—released Claude 4, a new family of large language models. For software development, it is said that the previous models produced errors in 30-40% of generated code. Claude 4 is rumoured to have dropped that to 6-10%. It certainly matches my own observations… This drop is not yet another incremental improvement. To me, it’s a phase change. When your error rate drops by that much, it fundamentally changes what’s possible. ...

Websites are dying. Not quickly, but they’re dying. The web has been a collection of pages for 30 years. You click, you scroll, you hunt for what you need. It’s slow and dumb and we’ve all just accepted it. But something’s shifting. Instead of browsing, we’re starting to talk to our computers. And they’re actually talking back. Why Websites Are Broken Here’s what you do when you want car insurance: Open Google. Click through 5-10 insurance sites. Fill out the same form repeatedly. Compare prices manually. Maybe use a comparison site that shows you 3 options instead of 300. ...

I finally got tired of pretending I’d one day set up a proper blog. This is it. I’ll write here occasionally — thoughts on systems, AI tooling, declarative platforms, and whatever else leaks out of my head. If you’re reading this, you’re either early… or lost. Either way, welcome. Who am I?