A job posting is a record of what someone believed they needed before the system was running. Ng assembled his four areas from more than 10,000 of them, plus interviews with experts, hiring managers and recruiters [2]. That method captures demand as employers articulated it, which is why the output looks like a build syllabus: LLMs, RAG, agentic workflows, architecture, testing, security, coding agents [3][4][5]. It is a survey of reqs, and reqs are written at the start of projects.
The arithmetic of the loudest complaint is off. One of the four items, "shaping the build," says engineers should no longer expect a pixel-perfect design to implement, and needs product sense plus an understanding of business context and customer goals [6]. So the reply calling the list "way too internally looking" and accusing it of failing to address the business problem [8] is quarrelling with a taxonomy that spends a quarter of itself on exactly that [15]. Problem framing is in there. What is not in there is anything about the second year.
That is where Andy Thurai of The Field CTO is more precise than the pile-on. His objection is temporal: the taxonomy covers Day 1 innovation, while in enterprise settings the hard part has moved to orchestrating, observing and paying for the thing, so the list optimises for the variable that matters least in production [10]. Set that against Ng's own claim that every developer, including data, DevOps and ML engineers, now needs AI engineering skills [7], and you have two people describing different halves of a lifecycle as the scarce one.
For anyone drafting a req this quarter, the usable contribution is Deepika Sidana's definition of orchestration: coordinating models, tools, data, evaluations, observability, human approvals and fallback paths [12]. Seven nouns [17], each of which becomes a line in a posting and a question with a checkable answer. Ask a candidate what their fallback path was when a tool call timed out in production. Ask who approved which class of action. Those are interviewable. Sidana's broader list of the business problem, customer workflow, risk, compliance requirements and the consequences of failure [11] is harder, because most of it is only demonstrable by someone who has already shipped and been on the hook when it broke.
The rest of the added vocabulary is weaker as hiring material. Deterministic governance, multi-agent orchestration, agent arbitration, AI FinOps, runtime economics and agentic observability [14] are terms a panel and a candidate would define differently on the same afternoon, which makes them unscreenable no matter how real the underlying work is. Cost ownership is the exception worth translating: someone has to be accountable for tokens per resolved ticket, and Ng's own framing already treats token waste as a competence question inside the coding-agent skill [5].
Sidana's closing condition is the one that does not belong in a skills matrix at all: accountability for measurable business outcomes [13]. That is scope and compensation, not a course.
