[danenania]

I'm working on an agent-based AI coding tool[1] that is trying to push the limits on the size/complexity of tasks that can be automated by LLMs, so I think about this often. I'm also using the full gamut of AI tools for development (I have 5 subscriptions and counting).

My opinion is that skilled engineers won't be replaced for a long, long time, if ever. While AI codegen will keep getting better and better, getting the best result out of it fundamentally requires knowing what to ask for—and beyond that, understanding exactly what is generated. This is because, depending on a particular prompt, there are hundreds/thousands/millions/billions of viable paths for fulfilling it. There are no "correct answers" in a large software project. Rather there is an endlessly branching web of micro-decisions with associated tradeoffs.

I think the job of engineers will gradually shift from writing code directly to navigating this web of tradeoffs.

1 - https://plandex.ai