AI Agents Are Coming: What Computer Scientists Need to Understand Beyond Chatbots

For many, artificial intelligence remains fundamentally transactional: a user poses a question, a system generates a response, and the exchange concludes. But a quieter and more consequential shift is underway. 

Emerging agentic AI systems represent a departure from content generation toward something closer to autonomous agency. Rather than waiting to be prompted, these systems can pursue defined goals, exercise judgment across a sequence of decisions, invoke external software tools, retrieve and synthesize information, and execute multi-step actions, all with minimal human direction.

That shift poses a foundational question for computer scientists: what changes when an AI system moves from recommending an action to executing it?

The Computer Science Foundations Behind Agentic AI

Understanding agentic AI requires engaging with computer science at every layer of the stack, and software engineering provides the scaffolding. Traditional software follows a predictable script: step one leads to step two, and the outcome is consistent. In agentic systems, the agent may approach the same task differently each time, encounter unexpected results mid-process, or need to recover from a partial failure. Engineers must design for that variability, building systems that can recognize when something has gone wrong, adjust course, and remain transparent about those adjustments.

Every agent needs memory to recall what it has already done and what it learned along the way. This is where data structures come in, and the implications extend well beyond technical performance. When an agent retrieves the wrong information or loses track of a prior step, the consequences can be significant and difficult to undo. For computer scientists, designing that memory correctly separates a system that can be trusted from one that cannot.

Databases and operating systems define the territory in which agents act. Unlike human users who query for information, agents interact with databases as autonomous actors with potential write access. Transaction management and rollback capabilities become critical when an agent can modify records at machine speed. Networking adds another dimension: most agents call external APIs, retrieve live data, and coordinate with other specialized systems. Every external connection is also a trust relationship that must be managed.

Security and verification are the disciplines most transformed by this shift. Prompt injection, in which malicious content in retrieved data can hijack an agent’s behavior, presents a distinct security challenge for autonomous AI systems. The principle of least privilege becomes more important when agents act autonomously. And because agent behavior is nondeterministic, traditional testing breaks down entirely, requiring behavioral evaluation, red teaming, and continuous production monitoring in its place. And because agent behavior can be nondeterministic, traditional testing becomes more challenging, requiring behavioral evaluation, red teaming, and continuous production monitoring.

The Computer Scientist’s Role in Agentic AI

Because computer science disciplines function differently in agentic AI, the role of the computer scientist must change as well.

Traditional development is prescriptive: a programmer specifies each instruction, and the system executes them in sequence. Agentic development is architectural. The computer scientist designs the environment, the constraints, and the interfaces within which an agent operates, then yields control over the specific path it takes. 

Boundary design is where this shift becomes most consequential. What permissions does an agent hold? What actions are prohibited? When must it pause and route a decision to a human? These are engineering decisions implemented in code. Logging and audit trails must be designed in from the start, and human-in-the-loop patterns are among the most important architectural choices in any agentic system.

The ethical dimension of agentic AI is critical. An agent acting autonomously can cause harm at a scale and speed that makes human detection difficult, and irreversible actions carry consequences that may outlast the task. Computer scientists who design agentic systems are designing accountability. Governance frameworks for these systems are still forming, and computer scientists are among the people shaping them. 

Preparing Computer Scientists for an Agentic Future

The rise of AI agents may appear to demand an entirely new form of computer science education; in practice, it reinforces the importance of many established areas of study.

Students who understand software engineering can design dependable systems around AI models. Knowledge of data structures helps them represent complex tasks and manage an agent’s decision process, while database courses prepare them to work with persistent data while protecting integrity and access. Operating systems and networking provide the foundation for controlling how agents execute, communicate, and use resources.

Data mining and machine learning give students a deeper understanding of how AI systems learn and make predictions, which is necessary for understanding when and why agents fail. AI-focused coursework can help them examine reasoning, planning, and intelligent system design, while security electives prepare them to recognize vulnerabilities created by autonomous tools.

At UoPeople, our computer science curriculum introduces students to these interconnected foundations. Just as importantly, the study of ethics and responsible AI can help future computer scientists examine the human consequences of the systems they create.

Specific tools and models will continue to change. The deeper skills — analyzing systems, identifying risks, evaluating evidence, designing reliable software, and making responsible decisions — will remain valuable across technologies.

Building Systems Worth Trusting

AI agents may change how people use computers. Instead of telling software exactly what to do at every step, users may increasingly describe a desired outcome and allow an agent to help determine how to achieve it. But greater autonomy increases the need for human expertise.

Computer scientists will be responsible for deciding what agents can do, what information they can access, how their actions are evaluated, and when human judgment must take precedence. Their challenge will not simply be to create systems capable of acting independently, but to create systems whose independence is appropriately limited, carefully monitored, and worthy of trust.