How big an AI multiplier an engineering team gets depends on the system its leader builds around it.
Read the essay →What actually changes when you build software with AI in the loop. One practice area per lesson.
I built an AI-native procurement tool end to end over several months, mostly on my own, with AI in the loop from the first plan through deployment. What follows is what I learned, the parts that held up and the parts that didn't.
Read it →The agent wrote the tests as it built the features, and they all passed, until a model change turned a pile of them red across the whole suite. Tests that cost almost nothing to write and a fortune to keep are a tax, and two rules fixed mine.
Read it →Two dependencies I lean on were compromised in the same week, days before I meant to call the build ready. I had deferred the hardening until the app was done, which is the plan that had just failed. Filing security under later does not work any more.
Read it →AI is faster than us on frameworks and patterns, but it optimizes for the thing in front of it, not the shape you'll need later. Here is how architecting a system changed once I was doing it with an agent, and where I learned to keep the judgment calls mine.
Read it →In a traditional app, QA validates repeatable behavior. When a model is doing the work, the behavior isn't repeatable, so validation has to live inside the app from day one. I learned that the hard way, minutes before a demo.
Read it →I wrote a lesson about validating a model's output, then a reader asked three questions about the validation layer itself that the post had no answer for. A gate nobody has watched reject anything isn't proven yet, and you can't make a model be wrong on demand.
Read it →