This is what I don't understand with AI: the prompt is a spec. But my experience is that writing a spec is terribly hard and tons of errors make it through.

Does anyone remember the old joke (what the client needed, what the spec requested, what the developer understood and what was shipped)?

With AI, who will write the prompts? Who will take the time to review and discuss them? If the past repeats itself, I foresee loads of inadequate software pushed onto unsuspecting users.

Dave Farley produced a recent video (3 weeks ago) where he states his opinion that the best use of AI is to give it narrow tasks, like one module at a time, and leave the big picture, like the system design, to a developer. That sounds reasonable. More a help completing the lengthy, boring part (coding) than autonomous programming.

How much does this mesh with the promises of agentic AI, where the agent is heralded as the autonomous employees replacement?

It might work on military projects, where the work is extremely divided and hierarchical. That is precisely the reason why I resisted working there (I did, tangentially: they were my customers for a few years).

In a world where war is increasingly fought by robots, it will be tempting to design these autonomous weapons using AI. Let's outsource war.. 💀