Pennings, A.J. (2026, Sep 08) From Computer Programming to AI Systems Engineering Management. apennings.com https://apennings.com/artificial-intelligence/from-computer-programming-to-ai-systems-engineering-management/

Introduction

Computer science is entering another major transition. For several decades, the central task of the computer scientist was to learn how to program computers. Algorithms, data structures, databases, operating systems, networks, and software engineering formed the intellectual and practical core of the discipline. Those skills remain critical. But artificial intelligence is adding another layer of computational engineering, one that changes what it means to build a computational system.[1]

The emerging computer scientist (and TSM graduate with CS specialization) increasingly needs to know how to construct systems that can retrieve information, connect to external resources, use computational tools, coordinate other AI systems, and act on the world. The important question is no longer simply, “Can you write the program?” It is increasingly, “Can you construct the computational system within which an AI model can operate?”

This distinction is important because artificial intelligence does not eliminate traditional computer science. It makes much of it infrastructure for something new. A retrieval-augmented generation (RAG) system still needs databases. An AI agent still needs APIs and networking. An MCP server still requires authentication, authorization, and software engineering. A multi-agent system requires state management, communication protocols, scheduling, and distributed-systems thinking. The traditional foundations remain, but they are increasingly assembled around a probabilistic computational component: the large language model (LLM).

The result is an AI systems layer on top of conventional computer science. This layer includes embeddings and vector databases, retrieval-augmented generation, large-language-model engineering, tool interfaces such as the Model Context Protocol, agent engineering, multi-agent orchestration, evaluation, AI security, human-AI workflows, and AI governance.

One of the most important new skills is understanding embeddings and vector databases. Traditional databases excel at exact queries. A vector database introduces another form of retrieval by representing information as numerical vectors and searching for items based on similarity in a high-dimensional space. A typical workflow begins with documents, divides them into chunks, transforms those chunks into embeddings, stores the embeddings in a vector database, and then retrieves semantically similar material when a query is submitted.

This changes the meaning of database engineering. Computer scientists no longer deal only with tables, records, fields, indexes, and SQL queries. The system must also deal with semantic representations, similarity metrics, vector indexes, metadata, chunking strategies, filtering, and retrieval performance. Databases increasingly become part of the knowledge infrastructure of an AI system.

This development is closely connected to retrieval-augmented generation, or RAG. The original RAG research proposed combining the knowledge contained in a language model’s parameters with an external, non-parametric memory represented through a dense vector index. The result was an architecture in which the model could retrieve relevant information rather than rely exclusively on what it encoded during training.

RAG therefore represents more than a technique for improving answers. It represents a new computational architecture. A production RAG system may involve query transformation, retrieval, metadata filtering, reranking, context construction, model inference, provenance, citation, and evaluation. Each stage presents an engineering problem. How should documents be divided? Which representations should be used? How many passages should be retrieved? Should keyword search be combined with vector search? Should results be reranked? How should contradictory sources be handled? How should the quality of retrieval be evaluated? These are as much computer science questions as they are AI questions. The model is only one component of the system.

The next major development is the ability to connect AI models to external computational resources. The Model Context Protocol, or MCP, is an open standard designed to connect AI applications with the systems where data and tools reside. MCP servers can reveal tools, resources, and prompts that AI applications can access. The significance of this development is conceptual as well as technical. The programmer is no longer merely writing an application that uses AI. The programmer can construct interfaces through which an AI system accesses the rest of the computational environment.

An MCP server might provide access to a database, document repository, Git repository, enterprise application, scientific dataset, financial information source, API, or computational tool. This makes designing machine-accessible interfaces an increasingly important computer science skill. Instead of designing only interfaces through which humans interact with software, developers increasingly have to design interfaces through which computational agents interact with software.

This development also raises familiar computer science problems concerning authentication, authorization, permissions, networking, reliability, and security. Giving an AI access to a database is not merely a matter of connecting an API. The system must determine what the AI is permitted to see, what it is permitted to change, how those actions are recorded, and what happens when an AI system makes an incorrect request.

The movement from models to agents adds another dimension. An LLM that answers a question is one thing. An AI system that can pursue a goal through multiple computational steps is something else. The basic architecture becomes a sequence of planning, tool selection, execution, inspection, revision, and verification.

The computer scientist must therefore manage state, memory, tools, permissions, error recovery, and workflow. The system must know not only how to generate a response but also what it is trying to accomplish, what resources it can use, what actions it has already taken, and whether those actions produced the desired result.

Multi-agent systems extend this architecture further. Instead of asking one AI agent to perform an entire task, developers can assign different computational roles to different agents. One agent might conduct research, another analyze the results, another write code, another critique the work, and another supervise the overall process.

This produces a new problem of computational coordination. The central question becomes how specialized agents communicate, delegate tasks, share state, resolve conflicts, recover from errors, and verify one another’s work.

In this respect, multi-agent AI brings an older computer science problem back into the center of AI engineering: distributed systems. Multi-agent architectures increasingly resemble distributed computational environments in which some components are probabilistic, autonomous, and capable of making decisions about their next actions.

This is why AI engineering should not be understood simply as a new form of programming or prompting. The programmer increasingly becomes an architect of computational environments. A traditional programmer might ask, “How do I write the program?” An AI systems engineer increasingly asks, “How do I construct the computational environment in which an AI system can retrieve information, reason, access tools, coordinate with other agents, and safely act?”

That question requires systems thinking. The engineer must determine where knowledge resides, what information should be retrieved, which model should perform a particular task, which tools the model can access, what permissions those tools should have, when another agent should take over, when a human should intervene, and how the entire process can be evaluated and audited.

The resulting architecture can be understood as a movement from data to retrieval, from retrieval to models, from models to tools, from tools to agents, and from agents to orchestration, human supervision, and infrastructure.

This represents a broader conception of computer science. Students still need algorithms and data structures. They still need databases, networks, operating systems, programming languages, computer architecture, and software engineering. But they increasingly need to understand how these technologies become infrastructure for artificial intelligence.

The computer scientist of the AI era therefore needs to be part programmer, part data engineer, part systems architect, part AI engineer, and increasingly part orchestration designer. The important competence is not necessarily mastering any particular framework. Frameworks will change rapidly. The more durable competence is understanding how computational components fit together into an operational system.

This may ultimately be the defining skill of AI-oriented computer science. It will be not merely writing code, but designing the computational infrastructure through which machines can know, retrieve, compute, coordinate, and act.

Traditional computer science has been exceptionally successful at teaching machines to compute. AI systems add increasingly important capabilities for retrieving information and taking action. The emerging computational sequence can therefore be expressed as Know -> Retrieve -> Compute -> Act.

That sequence suggests a significant change in how computer science education might be organized. AI should not replace computer science fundamentals. Rather, AI makes those fundamentals part of a larger systems architecture in which data, models, networks, databases, tools, agents, organizations, and people interact.

The deeper educational challenge is consequently not simply teaching students how to use artificial intelligence. It is teaching them how to engineer the computational systems through which artificial intelligence operates.

Notes

[1] This framing coincides with the Technological Systems Management (TSM) perspective and its Computer Science specializaiton where the emerging curriculum is not simply “computer science plus AI.” It is movement toward managing the architecture connecting data, models, networks, tools, agents, organizations, and human decision-making.

© ALL RIGHTS RESERVED

Not to be considered financial advice. AI is often used, and results are thoroughly interrogated. Links are used for some citations. Views are my own and do not express the stances of my employers, past or present.

---

Anthony J. Pennings, PhD is a Professor at the Department of Technology and Society, State University of New York, Korea and a Research Professor for Stony Brook University. He teaches AI and broadband policy. From 2002-2012 he taught digital economics and information systems management at New York University. He also taught in the Digital Media MBA at St. Edwards University in Austin, Texas, where he lives when not in Korea.

Related

Category: artificial intelligence
Tags: AI Systems Engineering

Comments

Comments are closed.