Career

Andrew Ng's AI Engineering Skills Map: The 4 Skills That Matter in 2026

Andrew Ng's AI Engineering Skills Map identifies four skills that separate effective AI engineers in 2026: building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build. The map draws on an analysis of 10,000+ job postings plus dozens of structured interviews with AI experts, hiring managers, and recruiters — Ng's data, presented as his synthesis. The September 4 deep-dive letter details the coding-agent skill, and its most quotable finding is a warning: the practical utility of very long-horizon agent runs “has been amplified beyond reality.”

The most important skills for using AI coding agents effectively — presenting the AI Engineering Skills Map for using coding agents.

— @AndrewYNg September 4, 2026

The Four Skills

SkillWhat it covers
1. Building & deploying AI applicationsLLMs, context engineering, RAG, agentic workflows, ML/DL — plus the statistical techniques to measure, steer, and govern unpredictable outputs. Core discipline: evals + error-analysis loops.
2. Software engineering fundamentalsTradeoffs (cost, scalability, reliability, speed, security/privacy); stack, architecture, data-store, testing. Lets you steer agents 'using the precise language of software engineering' instead of blind vibe-coding.
3. Using coding agentsMental model of how agents work, limitations + workarounds, context management, plan-vs-execute, verifiers/evals, spec discipline, multi-agent orchestration, safety (e.g. an agent wrecking a prod DB).
4. Shaping the buildAgents deliver to a spec, so engineer value shifts to deciding what is in the spec: product sense, business context, customer goals, ownership, MVP-vs-careful judgment.

A terminology note worth keeping: Ng says “AI Engineering skills” are not the “AI Engineer” job title — the analogy is that all developers now need cloud skills, while few hold a “Cloud Engineer” title. The map describes what every builder needs, not a niche role. A cross-cutting trait underlies all four: a continuous-learning mindset, because the tooling changes faster than curricula.

The Evidence Basis

Hedge this correctly: the basis is Ng's stated methodology — analysis of 10,000+ job postings, dozens of structured interviews with AI experts and hiring managers, surveys, and synthesis of other online data. The raw sample window, instruments, and clustering methodology are not published, so present the map as an expert synthesis, not an audited dataset. DeepLearning.AI keeps the map updated via an open community survey.

Using Coding Agents: The Deep Dive

The September 4 letter details the third skill. The workflow keeps the pre-agent software shape — planning (brainstorm, spec, execution plan, plan review), execution (calibrated autonomy, automated and human verification), deployment and monitoring (gated deploys, agents watching logs and proposing fixes) — but the engineer's attention shifts from writing code to specification quality and verification design. Greenfield specs can be loose prompts; brownfield work needs rigor. Five sub-skills, per the letter:

Sub-skillWhat it covers
Directing the workflowEffort allocation per step; when to iterate; speed/cost/risk/effort tradeoffs; architecture choice; spec detail; decomposition into verifiable steps; when to retain human ownership.
Enabling agent autonomyInteractive vs delegated vs loop-until-goal; context calibration across phases; parallel agents + orchestration; attention management; permissions and gating for safety.
Reviewing the workBehavioral + functional verification matched to task; user-flow tests with screenshots as evidence; eval sets and LLM-as-judge; agentic code review + AI security audits; selective human review.
Customizing agent + environmentSkills, plugins, MCP servers (pruned when obsolete); hooks; standing context (AGENTS.md/CLAUDE.md); cross-session state; retrospectives; clearing agent debt; team context coordination.
Coding-agent foundationsHow agents work: retrieval, context windows, tool/MCP context cost, agent-subagent interaction, harness-around-LLM; failure modes (overengineering, missing verification, stopping short, destructive actions).

Named tools in the letter: Claude Code, Codex, and Cursor as proprietary agents; OpenCode and Pi as open ones — with progress driven by both harness and model improvements, which is why the map includes a routine of re-testing tools as practices evolve.

The Long-Horizon Cost Warning

The letter's most valuable correction to social-media folklore: yes, sometimes it is useful to run agents autonomously for hours and burn millions of tokens — but “the practical utility of very long-horizon tasks, especially relative to their cost, has been amplified beyond reality.” Most effective agent use is a complex, highly iterative process, and high-skill human intervention beats heroic unattended runs. Oversimplified social-media descriptions are explicitly called out as unreliable.

Who Needs What

If you are a working developer: skills 2 and 3 are the near-term priority — fundamentals make you a better agent-steerer tomorrow. If you are a team lead or founder: skill 4 is your job now; the spec is where product and engineering judgment live. If you are ML-adjacent: skill 1 differentiates you — evals and error-analysis loops are the discipline most teams skip. And if you hire: the four skills make a cleaner rubric than “has used ChatGPT.”

A Four-Skill Hiring Rubric

Turn the map into interview probes (evidence basis is Ng's synthesis — calibrate to your stack):

  1. Building/deploying AI apps: “Walk me through an eval you built — what did error analysis change?” Strong answers describe loops, not one-off benchmarks.
  2. Software fundamentals: “Your agent just shipped a caching layer. What tradeoffs did it make?” Tests whether they can name reliability/security/cost tradeoffs the agent encoded.
  3. Using coding agents: “Show a task you delegate vs steer interactively — and why.” Strong answers reference verification design and context management, not tool loyalty.
  4. Shaping the build: “What did you cut from the spec, and what did that buy you?” Tests product judgment — the skill that survives model turnover.

FAQ

Is this map just marketing for DeepLearning.AI courses?

The map is freely published in The Batch letters and stands on its stated evidence basis; DeepLearning.AI sells courses, and skepticism about course-adjacent advice is healthy — but the four-skill decomposition matches what independent job-posting analyses and our own bench coverage (ETH Zurich's CS-knowledge finding) support.

Which of the four skills should I learn first?

If you already write software: skill 3 (using coding agents) pays back fastest, and skill 2 makes you better at it immediately. If you are new to software: fundamentals first — the map itself argues fundamentals let you steer agents in precise language rather than blind vibe-coding.

What is “shaping the build”?

The fourth skill: because agents deliver to a spec, the engineer's leverage moves into deciding what the spec contains — product sense, business context, and the judgment of when to build an MVP versus carefully. Ng teased a dedicated follow-up letter; check The Batch for it before citing specifics beyond the map.

Sources

Get the latest on AI, LLMs & developer tools

New MCP servers, model updates, and guides like this one — delivered weekly.