AI Engineering Skills Map

Source — Andrew Ng

Andrew Ng (@AndrewYNg) — DeepLearning.AI / Coursera co-founder.

Original post: x.com/i/status/2088302050706686198 (14 Aug 2026) — “The AI Engineering Skills Map.” Research base he cites: 10k+ job postings, dozens of structured interviews with experts/hiring managers/recruiters, surveys, online data. High bookmark signal (~13k bookmarks at capture).

He frames this as skills, not only the “AI Engineer” job title — same way all developers need cloud skills without being titled Cloud Engineer. Full-stack, data, DevOps, ML, and AI-titled roles all need this map.

Note: We already have /aiengineer (6-month technical roadmap from another source). This page is the Ng four-bucket map + how we use it on our own apps and agents.

One-sentence TL;DR

AI lets us build software differently than 2022. The four skills that cluster out of jobs + expert interviews are: (1) build & deploy AI apps, (2) software engineering fundamentals, (3) using coding agents, (4) shaping the build — underpinned by continuous learning. In practice we split work into software applications and AI agents & workflows, and both need the full map.

The map at a glance

Underlying mindset: continuous learning — AI changes fast; skills and workflows must keep evolving.

Bucket 1 — Building and deploying AI applications

Core difference vs classic software: outputs are unpredictable. Prompt an LLM or train a model and you don’t know the next token or prediction the way you know a deterministic function. So the skill is not only “call the API” — it’s measure, steer, and govern with statistical discipline.

Ng’s building blocks (as he lists them):

• LLMs and context engineering
• RAG
• Agentic workflows
• Machine learning and deep learning
• Evals and error-analysis loops (the operational heart)

How we do this in our apps

Bucket 2 — Software engineering fundamentals

Deep understanding of how software works → better tradeoffs (cost, scalability, reliability, speed) plus security and privacy. Without that vocabulary, vibe-coding hands the agent a vague goal and gets poor default tradeoffs — because you didn’t give context the agent needs.

Fundamentals let you steer coding agents with the precise language of software engineering (architecture, data stores, testing, stack choice).

How we do this in our apps

• Astro site (this repo) — routes, redirects, Netlify, content architecture; agents ship pages, humans own deploy risk and structure (Codex engineering, MCP vs CLI.
• Headless Domains / x402 / wallets — identity, payments, security boundaries are classic engineering tradeoffs with agent consumers.
• CLI-first tools — task-shaped interfaces agents can trust (/cli, /cli2.
• exe.dev / sandboxes — isolation, blast radius, production vs prototype machines /exedev.

Bucket 3 — Using coding agents

Agentic coding is now a core skill for every developer. That means a mental model of how agents work, their limits, when to intervene vs leave them alone, and how to avoid token waste and disasters (e.g. agent touching production data).

Ng calls out concrete sub-skills:

• Manage the agent’s context
• Trade off planning vs execution
• Help the agent close loops with verifiers / evals
• Work with a clear spec (and know when not to bother)
• Orchestrate multiple agents
• Avoid pitfalls (production DB, silent side effects)
• Routines to keep trying new tools as best practices change

How we do this in our workflows

Bucket 4 — Shaping the build

Given a clear spec, coding agents are getting better at delivering to it. So engineer work shifts toward what goes in the spec. No longer “here’s a pixel-perfect design, only implement.” You need product sense, business context, and customer goals — plus ownership and agency to pick problems and choose MVP vs careful build.

How we do this across applications vs agents

Two workstreams (how Mike sees it)

A simple overlay on Ng’s map — same four skills, two product surfaces:

You don’t pick one column. theflysales is an application that embeds agents; this blog is an application whose content is often written with coding agents; RankingSolution is application + agent workflow. The map keeps all four buckets honest.

Personal learning checklist (use the map)

1. Score yourself 1–5 on each of the four buckets (honest, not LinkedIn).
2. Pick the weakest bucket for the next 30 days — not a new shiny tool in a strong area.
3. Attach practice to a real app: one product feature (application) + one agent/workflow (fleet) so both columns get reps.
4. Require an eval or verifier on anything agentic (even a simple checklist).
5. Write the spec first when the agent will implement; write a field note when you’re shaping.
6. Keep a “try one new agent practice monthly” habit — Ng’s continuous learning bucket.

Related on this site

• AI Engineer 6-month roadmap — complementary technical curriculum
• AI engineer vs operator
• Codex engineering · Grok Build
• Multi-agent harness · God agent vs team
• Product brain · SEO loop · Company brain
• CLI for agents · MCP new · Harness tips
• 5 roles · Being proactive

Primary source: Andrew Ng — AI Engineering Skills Map · DeepLearning.AI will expand each skill in upcoming posts.