Andrew Ng's AI Engineering Skills Map: 4 Core Skills from 10K+ Job Analyses

Andrew Ng released the AI Engineering Skills Map in DeepLearning.AI's The Batch #366 (2026-08-14). The map is the result of a data-driven methodology: his team analyzed 10,000+ job descriptions , conducted dozens of structured interviews with AI experts, hiring managers, and recruiters, and layered on survey data and other online sources. Ng describes the process as clustering this massive dataset to surface the skills that matter not just today but in the near future.

A Clustered Skill List, Not a Curated Reading List

The clustering converged on four top-level capabilities:

Building and deploying AI applications — 构建与部署 AI 应用

Software engineering fundamentals — 软件工程基础

Using coding agents — 使用编程智能体

Shaping the build — 塑造构建方向

Each top-level skill branches into a hierarchy of sub-skills; the map is a capability tree, not four isolated topics.

Correcting the Label: AI Engineering Skills, Not "AI Engineer"

Ng stresses that the map covers AI Engineering skills , not the "AI Engineer" job title . The distinction is critical: just as every developer today needs cloud skills but only a few are titled "Cloud Engineer," every developer — full-stack, data, DevOps, ML, and AI engineers — needs AI engineering skills. This is a capability descent : AI engineering is moving from a specialist role to a universal baseline for all developers.

Today all developers must know how to use the cloud, but only a tiny fraction have the title "Cloud Engineer." Similarly, all developers — full-stack engineers, data engineers, DevOps, ML engineers, and of course AI engineers — need AI engineering skills .

What Problem Each Skill Solves

Skill 1: Building and Deploying AI Applications — Taming Unpredictability

The fundamental difference between AI software and traditional software is unpredictable output . An LLM prompt or a trained model yields non-deterministic results. Therefore, building AI systems relies heavily on iteration : skilled AI engineers repeatedly "build a small piece → check results → decide what to try next," with each step strongly influenced by intermediate outcomes.

Building reliable software systems on top of unreliable AI components — that is the core craft of AI engineering.

The key enabler is using statistical methods to measure, steer, and govern AI systems, especially running disciplined evaluation (evals) and error-analysis loops . This skill will be broken down into six sub-capabilities.

Skill 2: Software Engineering Fundamentals — More Important Because Agents Write Code

Counter-intuitively, Ng argues that because agents write code, software engineering fundamentals become more important, not less . Software engineering is about trade-offs among cost, scalability, reliability, speed, security, and privacy. Without the ability to see those trade-offs, a developer cannot make good decisions.

A novice relying on "vibe coding" may produce a simple app, but a coding agent will often make poor trade-offs across latency, availability, consistency, reliability, maintainability, simplicity, and cost — and the developer won't even know those trade-offs exist, let alone correct them.

Understanding software engineering fundamentals lets you command agents with precise engineering language instead of being led by them.

This skill expands into five modules: full-stack applications, data management, system architecture, security & reliability, and scaling & production operations.

Skill 3: Using Coding Agents — Mastery Means Steering, Not Just Calling

"Effective use of agentic coding is now a key skill for every developer," elevating coding agents from productivity tools to core competencies. Mastery means having a good mental model of how the agent works, knowing its limits and workarounds, and being able to steer it quickly — knowing when to intervene and when to let go — to produce reliable software without wasting time or tokens.

Concrete tasks include: managing agent context, balancing planning vs. execution, using verifiers or evals to help the agent self-correct, writing clear specs when needed (and knowing when not to), orchestrating multiple agents, and avoiding high-risk operations like accidentally dropping a production database. Because agentic coding evolves rapidly, mastery also requires a personal cadence of continuous experimentation to keep workflows aligned with emerging best practices. This skill breaks into five core capabilities.

Skill 4: Shaping the Build Direction — From Implementer to Definer

This is the most upstream and most overlooked skill. Ng observes that given a clear spec, coding agents' ability to deliver to spec is improving rapidly . Consequently, the engineer's center of gravity shifts upstream: from "implement the spec" to "decide what goes into the spec."

Engineers should no longer wait for a pixel-perfect design hand-off and then merely implement it.

Effective AI engineering demands product sense — understanding business context and customer goals — so engineers can participate in shaping, even leading, the build direction. AI also offers unprecedented ownership: proactively discovering valuable problems and responsibly delivering them. A key nuance is knowing when to ship a quick MVP for user validation versus when to slow down and refine. This skill decomposes into four key capabilities.

Not Four Parallel Courses: A Multiplication Problem

Treating the four skills as four equal-weight courses (additive) is wrong. The correct mental model is multiplicative :

Software engineering fundamentals = Foundation. If missing: agent makes poor trade-offs you can't see.

Building & deploying AI applications = AI system expertise. If missing: system runs but is unreliable and unmeasurable.

Using coding agents = Productivity lever. If missing: same model, others are multiples more efficient.

Shaping the build = Direction lever. If missing: you efficiently execute the wrong thing.

Multiplication means: if any factor is zero, the whole product is zero. You can maximize the productivity lever, but if direction is zero, output is zero — worse, you reach the wrong destination faster.

AI has not devalued foundational skills and product judgment. On the contrary, by driving execution cost toward zero, it exposes those two as the true bottlenecks.

As agent capabilities grow and the productivity lever equalizes across practitioners, the relative scarcity of software engineering fundamentals and shaping the build will only increase because they cannot be commoditized by tools.

The Underlying Mindset: Continuous Learning

The map concludes with a pervasive constraint: a mindset of continuous learning . At least two skills — using coding agents and shaping the build — see their best practices refresh every few months. You cannot "learn once"; you must build a routine mechanism for continuously trying new tools and evolving workflows .

AI is still changing rapidly, so each of us must keep learning and iterating our skills to embrace emerging best practices.

The personal takeaway: the most important change is not "take four more courses" but change the unit of learning . Yesterday's unit was "learn LangChain"; today's unit is "build a capability system that evolves with the frontier."

Reference: Andrew Ng, The AI Engineering Skills Map , DeepLearning.AI The Batch #366 (2026-08-14); X post @AndrewYNg (status/2088302050706686198). This article is a deep technical interpretation based on the original; the multiplicative framework is an analytical lens.