We might be already past the point where agents understand code better than the best developer, and they certainly are more efficient. We still want to feel like we're in control and have a grasp of what's going on, but it does seems like we are becoming the bottleneck. We're moving one abstraction layer up and only focusing on architecture/design, but I don't see any reason why architecture/design decisions won't also best left to agents.

Not sure what we'll be left doing: maybe agency itself (deciding to do something and prompting the agents to do it) will be the last thing to go?

I think we are already there, we need to decide which control we want to keep, but I think testing the product through main use cases and letting AI test edge cases already solve most of your issue, you can fix bugs faster than before so the approach of how code is delivered have to change

I think we’re not far away from that point. Maybe in a few years. I find as a dev I read my code less and less. I can still communicate with the agent by explaining the code it’s likely getting wrong, finding perf optimizations, or making architecture but I’m definitely reading it less than even 3 months ago