Generative AI and coding agents are rapidly changing software development, but DeepLearning.AI founder and AI scholar Andrew Ng says that does not mean developers can skip software engineering fundamentals.
Ng recently published the "Software Engineering Fundamentals" portion of his AI Engineering Skills Map. His point was direct: even if coding agents end up writing nearly all code in the future, developers still need to understand how software systems work. Without that, they cannot clearly tell AI what to build or judge which trade-offs are being made.
Vibe coding can produce an app fast, but that is not the hard part
Ng warned that beginners with no grounding in software engineering who build applications purely through vibe coding may be able to create simple products quickly. The problem is that a coding agent may make the wrong trade-offs across latency, availability, consistency, reliability, maintainability, simplicity and cost.
In some cases, the developer may not even realize those trade-offs exist in the first place.
Coding agents are widening the path to full-stack work
Ng grouped the software engineering skills that still matter in the AI era into five areas: building full-stack applications, managing data, designing system architecture, keeping systems secure and reliable, and deploying and operating production systems.
One visible shift, he said, is that coding agents are broadening developer roles. Engineers who used to focus on front end, mobile, or a specific stack can now use coding agents to handle work outside their usual area and gradually move toward full-stack development.
He did not present that as a reason to learn less. Ng said mature developers still need to understand the major parts of front-end and back-end systems, including UI components, caching, page rendering, API design, authentication, state and session management, asynchronous processing, data persistence, testing, security and accessibility. AI may help implement the pieces, but people still need to understand how the whole system fits together.
Data architecture may matter more than writing code
Ng also singled out data management as a core skill. His reasoning was that data architecture is often much harder to change than code. Even if AI agents can help with data migration later, choosing the wrong data model or storage design at the start can shape a product for a long time.
Developers need to understand a system's access patterns to decide what data should be stored, how long it should be retained, and whether the right fit is a relational table, document store, key-value system, graph model, or something else. Those choices directly affect speed, scalability, availability, reliability and cost.
He added that engineers still need to understand transactions, concurrency, data consistency, freshness, privacy, governance and compliance. In AI applications, that becomes even more important because the context used by an AI system may come from enterprise data sources. If the underlying data architecture is wrong at the start, the AI itself may not even "know what it doesn't know."
For that reason, Ng said building data infrastructure for AI agents, not only for traditional software or human users, will become a fast-evolving field.
System architecture still requires human judgment
System architecture is another area where Ng said coding agents cannot replace human decision-making. Before building a product, engineers need to define factors such as user volume, how important latency is, how important cost is, where the boundary between front end and back end should sit, where application state should live, and whether to use a monolith or microservices.
Even the choice of programming language, runtime, front-end framework, back-end framework and data technology involves trade-offs.
Ng said the right answer changes with the stage of a product. An architecture that works for a prototype may not fit the first production release, and an early production architecture may need another rebuild after user growth accelerates. Real engineering skill is not memorizing one "best" architecture. It is knowing what to choose at a given stage and when to change it.
That is where pure vibe coding runs into trouble most easily. An AI agent can generate a system that appears to work, and do it very quickly. But if the user does not understand trade-offs involving latency, availability, reliability or cost, it becomes hard to judge why the agent made certain architectural choices, and harder still to tell whether those choices were wrong.
Faster code generation raises the bar for security and reliability
Ng also said engineers in the AI era will increasingly need some of the skills associated with security engineers. Reliable systems require a full testing strategy, including unit tests, integration tests, test frameworks and coverage.
Engineers also need to know how to deal with failure. He pointed to practical questions such as what happens when an API hits a rate limit, whether a service can degrade gracefully after an outage, and how to reduce the blast radius of a single failure.
Security work, in his view, cannot be bolted on after a product is finished. Ng referred to the industry's "shift left" direction, which moves security work earlier in the development process. AI tools can already help scan code for vulnerabilities, check whether dependencies carry supply chain injection risk, and analyze the attack surface of cloud configurations.
Still, he said tools can help find issues, but developers need enough security knowledge to understand the issues themselves.
A demo is not the same as a production system
Ng listed scaling and operating in production as the fifth core skill, drawing a line that he sees as increasingly important in the AI era.
AI coding tools have made it easier than ever to build a demo. Serving real users is a different task. Engineers still need to understand the full software development lifecycle, including build and test, deployment environments, release strategy, CI/CD, automated deployment and infrastructure as a service.
Once a system is live, another set of issues appears: observability, alerts, incident management and scaling. As traffic grows, developers need to know how to scale servers, load balance, and adjust data systems through sharding, indexing, replication or architectural changes.
Version control, code review, dependency maintenance and technical debt management also remain central to whether a product can keep evolving over time.
Ng's conclusion was clear in substance: AI can make the jump from 0 to 1 very fast, but the difficult move from 1 to 100 still depends heavily on engineering judgment.