Skip to content
Andrew Ng Reveals Four AI Engineer Skills That Coding Agents Can't Replace
Industry Insights 13 min read

Andrew Ng Reveals Four AI Engineer Skills That Coding Agents Can't Replace

Andrew Ng outlines four critical AI engineer capabilities—building and deploying AI applications, software engineering fundamentals, effectively using coding agents, and participating in product decisions—emphasizing that human judgment remains essential for trade-offs and business context that agents cannot grasp.

macrozheng
macrozheng
macrozheng

Code Verification Advances, But Enterprise Decisions Lack Standard Answers

At Neo4j's NODES 2026, CEO Emil Eifrem interviewed Andrew Ng about what AI engineers still need to learn given rapid coding agent progress. Ng's answer was not a new tool, but a capability map derived from analyzing over 20,000 job descriptions plus interviews and surveys.

Ng: AI excels at tasks with verifiable outcomes like programming and math. Judgment problems usually have multiple good and bad answers; we struggle to build equally clear reward and test environments for them.

Ng links the AI engineer shortage to this gap: coding agents accelerate development, yet teams still need people to judge whether requirements, data, and results align with business goals. He notes the "shortage" is his observational assessment, not a statistical vacancy count from the job descriptions.

A Capability Map: Four Interlocking Skill Groups

Ng describes four capability groups that an application cycles through from idea to production—not four isolated exams.

Ng: The four groups are building and deploying AI applications, software engineering fundamentals, using coding agents, and participating in deciding what to build. People skilled in AI engineering don't just implement specs; they also drive the development process.

1. Building and Deploying AI Applications

Turning a working prototype into a continuously running product. Models must integrate with tools and data, errors must be detectable, and runtime behavior must be monitored.

2. Software Engineering Fundamentals

Determines whether the product remains stable. Developers still need to understand system architecture, interfaces, testing, and security; these concerns do not disappear because an agent generated the code.

3. Using Coding Agents

Not merely letting agents write more code. Developers must articulate goals and context clearly, understand what changes the agent made, and verify that results meet requirements.

4. Participating in Deciding What to Build

Occurs before coding starts and continues throughout: defining the user problem, prioritizing requirements, and setting boundaries for AI autonomy. Without clarity here, faster development may only produce an unwanted product faster.

Using Agents Requires Understanding Their Trade-offs

When asked why developers still need computer science fundamentals if agents write code, Ng's answer centered on "trade-offs."

Ng: Not understanding fundamentals while having a coding agent write programs sometimes equals letting it choose for you among cost, latency, reliability, and security—without you even knowing a choice exists.

Agent-generated code may run and pass tests yet be unsuitable for production. Examples: cache TTL set too long serves stale data; over-permissive interface permissions expose sensitive information; mishandled concurrency causes duplicate operations. Developers must grasp these issues to spot the agent's implicit choices and decide what needs correction.

Ng does not expect developers to memorize all syntax—he himself writes less boilerplate thanks to agents. The difference: for a low-stakes one-page tool, an agent's arbitrary viable solution may suffice; for enterprise applications with strict permission and reliability demands, a developer who cannot see the underlying choices cannot steer the agent toward a suitable implementation.

Models Read Fields But Miss Business Meaning

The conversation turned to "ontology"—a clear definition of business concepts and their relationships: does "customer" mean individual or company? How do orders link to contracts? Do identically named "status" fields across systems share semantics?

Without this clarity, agents may misinterpret fields or join unrelated data. Ng emphasizes that application builders hold vast context the model never receives. He illustrates: a customer's facial reaction during a chat may reveal feature popularity; such signals rarely enter requirement documents but shape product and data organization.

Ng: Application builders possess massive context that coding agents lack. We know our business, where data comes from, which datasets are reliable, and we may even forget to tell the agent.

Ng does not oppose AI-assisted ontology construction—his team researches it—but observes that fully automatic generation currently underperforms human-guided approaches. In data science projects, he has seen AI reason from raw data to a visibly wrong assumption; correcting the high-level concept first made results sensible.

Ng: Automating this structuring is worth researching, but human-in-the-loop guidance still yields better results. AI sometimes makes substantive erroneous assumptions about data.

Making Enterprise Data Agent-Ready Is Not a Single-Vendor Data Dump

The host framed the web of people, projects, data, and business rules as an "enterprise world model." Ng addressed a common large-enterprise ask: how to adapt existing data architectures—built for humans and analysts—so agents can use them.

Ng: Making enterprise data suitable for agents involves permissions, governance, security, and the entire organizational data architecture. Graph, structured, and unstructured data, plus transactional system data, often need joint consideration.

Ng rejects the idea of handing all data to one vendor. Large organizations store orders, contracts, and support logs in separate systems with distinct access controls. Instead of "dump everything into one lake," the priority is clarifying provenance, access rights, and audit trails for each dataset.

Ng: Every data vendor wants to say "give me all your data, I'll clean it," but large enterprises won't hand everything to a single vendor. We must figure out how different data architectures combine.

From his project experience, Ng notes that early phases often consume massive time cleaning and connecting data; once the data foundation stabilizes, subsequent projects accelerate. This is his anecdotal observation, not a universal efficiency metric.

Closing Example: A Customer-Service Agent That Refunds Orders

Code that compiles and tests that pass cover only part of the work. Before launch, the team must still confirm: who defines refund rules, which system owns order status, what customer data the agent may access, and which scenarios require human escalation. When errors occur, traceability to the source data is essential.

Ng's four capability groups gain weight in such a project: one person drives development with the agent, another spots engineering trade-offs, another integrates the AI into real services, and another defines what the agent is allowed to do on users' behalf. The AI engineer's job has not vanished with faster code generation; the difficulty has shifted to the places where multiple people must jointly specify and verify.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

software engineeringAI deploymentAI engineeringdata architectureEnterprise AIontologyAndrew Ngcoding agentsAI skillshuman judgment
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.