Coding agents are good, but once the complexity is high they're not good enough. Eventually your agent won't be able to make changes to your code base without introducing bugs with every change. In my experience, the agents aren't very good with abstractions yet, and no amount of testing can completely paper over that problem. So yes, the industry is changing dramatically and at a breathtaking rate, but I don't think money is the only moat left.
Not just complexity, but also anything requiring any actual ingenuity. It's impressive how "junior" it can be, with its (current) statistical shackles. Worse, if you flat tell it what to do, if your approach is too far into the statistical weeds, it'll flounder like crazy.