Iâm a heavy AI user these days â most of my commits are co-authored with Claude. My output is much higher than it would be otherwise, and Iâm having more fun.
Yet in my experience, it doesnât pay off to follow the bleeding edge of agentic engineering. I donât doubt that some people get value from Ralph loops or Superpowers, but I canât help but notice the parallels to productivity influencers who treat optimizing your system as an end in itself.
This isnât to say that there isnât a lot to learn at the bleeding edge! But it is, by its nature, highly experimental and a lot of the methods wonât stick. We can be thankful to the people writing speculative fiction like Gas Town, but most of us donât need to think about it in our day-to-day.
For most engineers, I recommend the Normie Loop: use Claude Code or Codex or Conductor or whatever, learn to use it well, and talk to the model to steer it as needed. As these tools evolve in response to emerging (but validated) best practices, continue to evolve with them. This is where you get the best ROI.
The most useful (and measured) resource Iâve seen so far is Simon Willisonâs Agentic Engineering Patterns project.