We've Passed Another Inflexion Point

Last December (2025), we reached an inflexion point. That was the point at which there was no need to write code yourself. I believe in the past couple of months we’ve reached a new inflexion point. The point at which there is no need to read code yourself.

Because these inflexion points are such big changes, you have to force yourself to change, and as such I’ve set a goal for myself to stop reading code.

The best way to feel confident about the code you’re shipping is to build verification mechanisms. Let me demonstrate with the most recent example I’ve experienced.

For the past few years I’ve been working on an app that writes workout programs as well as or better than a personal trainer. My results have been subpar, and I think that’s because I felt the need to read the code and verify the prompts and business logic to understand why it produced what it produced. It was very tiresome, and I’ve never prioritised this project.

Fast-forward to today. Instead of reading the code, I ask my coding agent to build a playground, in this case a website, that shows me everything I care about. I can see what data is being passed around, the exact system prompt, the output, and whatever else I decide to add. The app improves by me looking at that data and talking to my agent about what it’s getting wrong.

Now my iteration speed is incredibly fast, and I’ve never been able to generate such high-quality programs. It won’t be long before these programs are as good as or better than a personal trainer.

One concern you might have with this approach is how you know the verification mechanisms are working as expected. Well, that type of problem has always existed. How do you test your tests? You can’t. You have to trust that they’re working as expected, and that’s easier to do because you wrote the tests, and tests are usually easier to write, and therefore verify, than the code for the core product. This is the same for the verification mechanisms you build for your coding agent. We’re able to read and view these mechanisms ourselves, which is a lot easier than reading code.

The fact is, verification mechanisms don’t catch everything. But in a land of coding agents, they enable autonomous work and potentially self-improvement, which I’m willing to bet does a better job at building production-ready software than manually writing tests or manually reading code. This makes for something a lot more robust and capable than in our prior times.