Nice article, putting names onto a wide spread symptom experience by a group of people.

Myself have a few realizations that I myself find very revealing of my own mental gap/blindness in retrospective:

1. The majority of socalled software engineers are destroyed/exposed by agentic coding tool. Among the reasons, one of them is that they are now addicted to disruption. I realized this through my own work where I utterly destroyed my own sense of software structure by overseeing close to 10 coding sessions across cursor cloud agents and multiple concurrent ones running on desktop. And, I noticed one of my team member becomes not able to make progress even he told me the clear plan, and during one in office work session, I noticed that he habitually browses his phone between cursor working sessions. And at that moment, I realized that person is no longer capable of solving nontrivial problems because his focus is utterly destroyed by the deprived thinking process by cursor.

2. Using AI coding agents to build conventional software is only meaningful for the top team. Put in other words, AI coding agents can only automate the parts of conventional software that is mundane and boring. Note that I am talking of building conventional software, not conventional software engineering. The conventional software engineering best practices still apply to agentic software engineering. Agentic software engineering appears necessary and fitting to building agents themselves, because that's guardrails are crucial, and manually putting guardrails are not a match to agents novel behaviors. And conventional software engineering really is not useful when reasoning and designing agents.

3. Speculatively, I am guessing we'll have some chances of really making breakthrough in designing an agentic safe programming language, which both serve as the constructs of ensuring alignment/safety/etc, and also allows reasonably simple understandability for humans to understand agents output. For one thing, it really makes sense to me that coding agents model CoT should be in the form of some semi/fully-formal (thinking of a better uml, but I don't know what shape it should be) languages that is more effective and precise than natural language, but is also machine verifiable.

4. It's very easy to get used to something that is reliably meeting my expectations. I was shocked by first sight of the unrecognizable but "working" code generated by Claude code by a team member. And when I tried to write an end to end feature myself on platform that I had no experience before, I suddenly become comfortable not looking at the code at all. The transition is as instinctively as how shocked I was at the beginning. And then it takes a long and conscious looking back to dial back the addiction to see that blindness and recover from that.