Table of Contents
Key Takeaway
- 🗺️ The map is official: Andrew Ng and DeepLearning.AI published the AI engineering skills map — a structured breakdown of the four highest-level skills every AI engineer needs in 2026, laid out in a series of public letters.
- 🧩 The four skills: Building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build — ordered from what you ship to how you steer the work.
- 📐 The shift from 2022: AI now lets engineers build software differently — the people who learn the new workflow first get, in Ng’s words, “numerous exciting project and job opportunities.”
- 🎯 The play: A 90-day plan mapped to the four tiers — build a deployed app, cement fundamentals, make agents daily tools, and practice shaping work — turns the map from reading material into a career move.
Andrew Ng just answered the question every ambitious professional is asking: with AI writing more of the code, what exactly should you be learning? His answer is not a course bundle — it is a map. The AI engineering skills framework, published by DeepLearning.AI as a series of letters from Ng, breaks the profession into four top-level skills and then details each one: building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build. The framing matters because it comes from the person whose courses trained a generation of machine-learning engineers — and because it describes what employers now hire for, not what was true in 2022. Ng’s own framing: AI lets engineers “build software very differently today than in 2022, and everyone with the skills to take advantage of this shift has numerous exciting project and job opportunities.”
What the AI Engineering Skills Map Contains
The map’s top level is deliberately small: four skills, not forty. In Ng’s presentation, these are the highest-level capabilities — (i) building and deploying AI applications, (ii) software engineering fundamentals, (iii) using coding agents, and (iv) shaping the build — with each broken into sub-skills in follow-up letters. The structure reads like an O*NET occupation profile for a job title that did not exist five years ago, which is fitting: the same classification thinking that powers labor statistics has been adapted for AI work by Epoch AI’s research taxonomy, and OpenAI used that taxonomy in September to classify how its own researchers delegate to agents.
The first skill — building and deploying AI applications — is the one Ng fleshed out first, and it is deliberately end-to-end. Not “learn machine learning theory,” but: take an application from idea to users, including the LLM integration, context engineering, retrieval, evaluation, and the operational discipline of keeping it working after launch. The second letter in the series defends the unglamorous half: software engineering fundamentals remain essential precisely because agentic coding tools make it easier to generate code than to judge it. Engineers who cannot review, test, and architect what agents produce become bottlenecks to their own productivity.
The third skill is the one that did not exist as a job requirement two years ago: using coding agents. Ng’s letters treat this as a distinct competence — knowing what to delegate, how to write the task so an agent succeeds, when to intervene, and how to run multiple agents in parallel, the specification discipline that underpins agentic security as well, as our prompt-injection defense guide set out. The fourth, shaping the build, is the judgment layer: deciding what to build, scoping requirements, and steering technical direction. As agents automate more of the typing, that skill moves from senior-engineer luxury to the core of the role — the same pattern OpenAI’s internal data showed this month, where researchers who manage fleets of concurrent agents out-produce serial users roughly three to one.
Why the Map Matters More Than Another Course List
The skills market for AI is loud, and most of the noise is generic: “learn AI,” “take a prompt engineering course,” “future-proof your career.” The AI engineering skills map cuts through because it comes from someone who has watched hiring from both sides — Ng founded Coursera, founded DeepLearning.AI, teaches the most-enrolled AI courses in the world, and reads thousands of job postings through his community. When his framework says the profession is now organized around four skills, hiring managers can map their job postings onto it, and candidates can audit themselves against it in an afternoon.
It also gives structure to a hiring pattern the industry has been reporting all year: the rise of the associate AI engineer. Skills-map analyses describe a ladder — associates who hold core skills and are actively developing expanded ones, full AI engineers who own core skills and are fluent in two or more areas. That ladder matters because it says the door is still open: you do not need to arrive complete. You need to arrive building. The letters’ career framing is blunt about who wins: “everyone with the skills to take advantage of this shift has numerous exciting project and job opportunities” — and the shift, unlike past platform changes, is one individuals can adopt without permission from an employer.
There is a generalist dividend too. DeepLearning.AI’s letters have repeatedly argued that AI rewards generalists — full-stack engineers joined by full-stack recruiters and full-cycle marketers, because AI assistance collapses the specialization barriers that used to make solo end-to-end work impossible. The AI engineering skills map is the formal version of that argument: the four skills compose, and professionals who assemble all four become the scarce generalists the market is hiring.
The Filipino and Global Professional Angle
For Filipino professionals — and for the global professionals reading alongside them — the map lands at a useful moment. The remote-work market that Filipino engineers, analysts, and creatives serve is being reorganized around exactly these four skills, and time-zone-arbitrage careers now compete on output rather than hours. An engineer in Manila or Cebu who can deploy a working AI application, keep it maintained, drive coding agents through a full workday, and scope the work herself competes at the same level as anyone in any timezone — the tools do not care where you sit. Our coverage of the AI jobs debate among Hinton, Li, and Ng traced how the economists and practitioners converge on exactly this: leverage comes from building with AI, not competing against it.
The local evidence supports the map’s shape. The Philippines’ IT-BPM sector — the country’s flagship services industry — is repositioning around exactly the higher-judgment work the map describes, and the professionals moving first are the ones treating agentic tools as daily infrastructure. The same reasoning runs through our AI cybersecurity career analysis: the durable roles are the ones that direct, audit, and secure the machines.
The 90-Day Plan, Skill by Skill
The map becomes a career move when it gets a calendar. Here is the 90-day conversion of the AI engineering skills framework, one skill per three weeks, with the evidence that proves it worked.
Days 1-21 — Building and deploying AI applications. Ship one deployed application with real users, however small: an internal tool, a public utility, an API-backed workflow. The point is the full loop — requirements, build, deploy, monitor, iterate. Evidence of completion: a URL someone other than you has used, plus a note of what broke in production and what you changed.
Days 22-42 — Software engineering fundamentals. Pick the fundamentals gap that hurts most — testing, code review, system design, or git discipline — and apply it to the app you just shipped. Write tests for its core paths, put every change through review, document its architecture in one page. Evidence: a repository that a stranger could extend without asking you anything.
Days 43-63 — Using coding agents. Make agents the default, not the garnish: every coding task starts as an agent delegation, and your job is specification, review, and integration. Run two agents in parallel on independent tasks from week one, and measure the intervention rate — how many tasks needed your correction. Evidence: a week where agents produced more committed code than you typed, with your review as the quality gate.
Days 64-90 — Shaping the build. Practice the senior layer: write a one-page spec for a feature, get it scoped and estimated, decide what to cut, and direct the agents through it. Volunteer for the prioritization conversation at work — the one that decides what gets built. Evidence: a shipped feature whose scope you defined, not just whose code you wrote.
At the end of the 90 days, the portfolio says what no certificate can: four skills, each with shipped evidence. That AI engineering skills portfolio is the audit the skills map implies — and it maps directly onto the job postings that now read “AI engineer (associate)” and mean it.
What the Map Gets Right — and What It Leaves to You
The honest reading of the skills map includes its gaps. It is a competence framework, not a job guarantee: the market still rewards the framework’s skills unevenly by region and industry, and “numerous exciting opportunities” in aggregate can still feel scarce from any single job search. It also presumes the time to build — 90 days of deliberate practice is a real cost for professionals working full schedules, and the letters’ optimism is aimed at people who can carve that out.
What the map leaves to you is sequencing and evidence. The AI engineering skills are equally ranked at the top level, but careers are not: an engineer with deep fundamentals but no deployment experience reads differently from a builder who cannot review agent output. The self-audit question the map enables — which of the four is my weakest? — is the whole point, and the honest answer usually points at the next 90 days rather than the last certificate earned.
Tools and Resources
Andrew Ng’s letters — The Batch, AI Careers collection — the primary source for the skills map and its detailed skill-by-skill letters.
DeepLearning.AI — Ng’s education company and the home of the AI engineering curriculum the map organizes.
WEF Future of Jobs Report — the employer-side dataset on which skills companies expect to hire for through the decade.
Frequently Asked Questions About the AI Engineering Skills Map
What is the AI engineering skills map?
It is Andrew Ng and DeepLearning.AI’s structured framework for the AI engineering skills professionals need in 2026, published as a series of letters. The top level contains four skills: building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build — each broken into sub-skills in follow-up letters.
What are the four highest-level AI engineering skills?
Building and deploying AI applications; software engineering fundamentals; using coding agents; and shaping the build. Ng describes these as the highest-level skills, each expandable into detailed sub-competencies in the accompanying letters.
Do I need a computer science degree to follow the map?
No. The map is organized around demonstrable skills rather than credentials, and DeepLearning.AI’s letters emphasize that everyone who builds the skills — including people entering from other fields — has project and job opportunities. Evidence of shipped work matters more than a specific degree.
How long does it take to build these skills?
A disciplined 90-day plan covers the loop: ship a deployed AI application in month one, strengthen fundamentals against it in month two, make coding agents your daily workflow in month three, and practice scoping and directing builds throughout. The timeline assumes evenings and weekends around a full-time job.
Is “using coding agents” really a distinct skill?
Yes — and it is the newest one. Delegating effectively to agents requires writing well-defined tasks, judging output quality, intervening at the right moments, and running parallel sessions. OpenAI’s own research-organization data shows the productivity difference between serial and parallel agent use is roughly threefold at the frontier — the same agent-workday ratio behind the automated research intern milestone.
How does the map help people outside the US?
The skills are location-neutral and verifiable through public artifacts — repositories, deployed apps, written specs. Remote hiring increasingly screens for exactly this evidence, which lets professionals in any market compete on output. The framework’s global accessibility is the point: the map tells you what to build regardless of where you build it.
Financial Disclaimer
This article is for informational and educational purposes only and does not constitute professional career or financial advice. Program terms and market conditions change; verify current details with each provider before making education or career decisions.
