In this X article, Jacob Sansbury, founder of Spawn, describes how Spawn is run entirely by agents and how they spend monthly 30 times more on tokens than on payroll.
At one point, Jacob talks about the fact that it's closing the loop that allows agents to be so effective at working on code bases even though they also create bugs and break things. This got me thinking.
If you develop at Human speed, perhaps with some agentic augmentation, then things work. You may be a bit faster because an LLM gives you a push here and there, but if it's mostly Humans doing most of the work, you're in the realm of what software engineers have been doing for ages.
If you develop and review at agentic speed, but then need Humans in the loop to review bug reports or other breakages, the system simply doesn't work. You're creating changes at a pace that no Human can keep up with, and you'll get to a point where you have a codebase that you don't understand and that you can't work on to fix the problems that arise.
If you develop and review at agentic speed, you need to have agents dealing with the bugs that other agents created. If you do this, then the system might work.