I’m the author. This article began with a question I keep returning to: what separates legitimate AI-assisted programming from “vibe coding”?
To me, the difference is understanding the fundamentals, making informed low-level technical and architectural choices, and being able to diagnose and repair the system when it fails under real constraints with real users depending on it. “It works” is not enough if you do not know why, how reliably, or whether it is the right solution.
AI can also create enormous working-memory demands. It lets individuals build more complicated systems, while team coordination and the underlying hardware and software stack continue to grow more complex.
I’m not convinced by fully agentic programming. It looks to me like an exquisitely complicated Rube Goldberg machine for what was traditional in the realm of batch programming. I see the more promising end state as cognitive pair programming: a human programmer working closely with an LLM, while retaining control. In fact, I would say Agentic AI as I seen it proselytized is more about certain LLMs trying to eat as much engineering budget as possible, and is the real threat to jobs, and intent, quality, etc.
I’d welcome any thoughts, especially disagreement.
2 comments
[ 3.7 ms ] story [ 4.0 ms ] threadTo me, the difference is understanding the fundamentals, making informed low-level technical and architectural choices, and being able to diagnose and repair the system when it fails under real constraints with real users depending on it. “It works” is not enough if you do not know why, how reliably, or whether it is the right solution.
AI can also create enormous working-memory demands. It lets individuals build more complicated systems, while team coordination and the underlying hardware and software stack continue to grow more complex.
I’m not convinced by fully agentic programming. It looks to me like an exquisitely complicated Rube Goldberg machine for what was traditional in the realm of batch programming. I see the more promising end state as cognitive pair programming: a human programmer working closely with an LLM, while retaining control. In fact, I would say Agentic AI as I seen it proselytized is more about certain LLMs trying to eat as much engineering budget as possible, and is the real threat to jobs, and intent, quality, etc.
I’d welcome any thoughts, especially disagreement.