Good points and questions and I don't have the answers.

I've definitely thought about how I currently offer something to the company because I know these systems and can stop the agents from doing weird shit. But in a greenfield project I would probably (definitely) not get the same understanding.

I think as we figure out best practices - you'll start to see that you can rely less on humans. More tests, more acceptance tests of critical business logic, etc.

I think we might get to the point where I'm not really adding value anymore because I dont know the codebase well enough to stop the agent. But then again - the cost of code is getting cheaper and when you reach that state where agents can't reliably work in the codebase due to AI slop - maybe you just start over? Kick off a few agents over the weekend and come back on Monday to a new (hopefully leaner) codebase? Run it in shadow mode next to the production service - ask the agents to fix any discrepancies and iterate until you have a simpler service?

I dont know what will happen. But I dont think we'll go back