Types of AI: Narrow, General, and Where Agentic AI Fits

Headlines often talk about “AI” as if it were one thing, from the spam filter in your inbox to science-fiction machines that think like people. In reality these sit at very different levels of ability, and mixing them up leads to muddled thinking, and to wrong answers on exams. This post sorts AI into its main types by capability, shows how the common terms fit together, and explains where agentic AI, the topic of Part D, belongs in the picture.

1. Two different questions

Most confusion comes from mixing up two separate ways of describing an AI system:

• How capable or general is it? This gives the familiar scale of narrow AI, general AI and superintelligence.
• How is it used? This describes a system’s role: a classifier that labels data, an assistant that answers questions, or an agent that pursues a goal by taking actions.

Agentic AI belongs to the second question, not the first. Keep the two separate and the rest of this post falls into place.

2. Types of AI by capability

Narrow AI (artificial narrow intelligence)

Narrow AI is built to do a specific task, or a limited set of related tasks, and cannot step outside that scope on its own. A model that detects fraudulent card payments cannot translate a document, and a chess engine cannot recommend a film. Every AI system deployed today is, in a technical sense, narrow, even though some are impressive.

• Everyday examples: spam filters, product recommendations, face unlocking, speech-to-text, route planning, fraud detection.
• The philosophical cousin: Searle (1980) drew a line between “weak AI”, where a computer is a useful tool for studying or imitating the mind, and “strong AI”, where a correctly programmed computer would genuinely have a mind. Narrow AI is an engineering term about scope, while weak and strong AI are a philosophical argument about minds, so avoid treating them as the same pair.

Artificial general intelligence (AGI)

AGI means a system with human-level flexibility: it could learn and perform a wide range of intellectual tasks, transfer what it learns to new situations, and cope with problems it was never specifically designed for. The term gained currency in the 2000s through the research community around that goal (Goertzel & Pennachin, 2007).

• It remains a goal, not an established fact. There is no agreed definition and no agreed test. Legg and Hutter (2007), for example, collected many definitions of intelligence and proposed a formal one based on performing well across a wide range of environments.
• Foundation models blur the picture. Large models handle far more tasks than older narrow systems, and people disagree about how to describe them. Most practitioners would not call current models AGI, but the question is genuinely debated. For your exams, treat deployed AI as narrow unless a question states otherwise.

Artificial superintelligence (ASI)

ASI is a hypothetical system that would greatly exceed human ability across almost every field. Bostrom (2014) examines how such systems might arise and what risks they could pose. It is a topic for debate, not a product you will configure on AWS.

Summary

Type	What it means	Status today	Example
Narrow AI	Excels at a defined task or set of tasks	Exists and is widely deployed	Fraud detection, speech recognition, recommendations
General AI (AGI)	Human-level flexibility across many tasks	Hypothetical; a research goal	A system that learns any new job as a person can
Superintelligence (ASI)	Far beyond human ability in nearly every area	Hypothetical; a subject of debate	Appears in discussions of long-term AI risk

3. Types of AI by what they produce or how they work

For practical work, you will use other labels far more than the capability scale. These are the ones that matter for the exams and for building real systems:

• Rule-based (symbolic) AI vs learning-based AI: explicit IF–THEN rules versus models that learn from data. This is the split you saw in the history post.
• Predictive (discriminative) AI vs generative AI: models that classify or predict versus models that create new content.
• Supervised, unsupervised and reinforcement learning: how a learning-based system learns. These each get their own post later in Part A.

4. Where agentic AI fits

An agent is a long-standing idea in AI. Russell and Norvig (2021) describe an agent as something that perceives its environment and acts on it to achieve goals, and they organise the whole field around this idea. What has changed recently is that large language models now make it practical to build agents that work with open-ended language, tools and software.

Agentic AI describes systems that are given a goal and then work toward it with some autonomy. They typically plan steps, use tools such as search or software interfaces, observe the results and adjust, repeating this loop until the job is done or a person steps in. Terminology in this area is still settling, so expect different sources to draw the boundaries differently.

• It is a way of using AI, not a higher level of intelligence. An agent can be built from narrow components, including a foundation model, and does not need to be AGI.
• Autonomy is not the same as intelligence. An agent may act on its own and still fail at something a child finds easy, so oversight and guardrails remain essential.
• It builds on what you already know. Agents combine foundation models (Part B), retrieval (Part C) and tools, which is why Part D comes after those parts.

Assistant, workflow and agent

Pattern	Who decides the steps?	Example
Chatbot or assistant	The person, one request at a time; the model answers	Asking a model to summarise a document
Workflow	The developer, in a fixed sequence of steps that may include a model	Extract text from a form, then classify it, then store the result
Agent	The system, within limits set by the developer, choosing steps and tools as it goes	Researching a topic, running tools and compiling a report

Part D explains when each pattern is the right choice and how the AWS agent services work. For now, remember that a fixed workflow is often cheaper, faster and more predictable than an agent, so “use an agent” is not automatically the best answer.

5. Common confusions and exam traps

• “Narrow” does not mean “simple” or “weak”. A narrow system can be extremely capable at its task.
• Agentic AI is not AGI. One describes how a system is used, the other describes how general its intelligence is.
• Foundation models are broader than older models, but that does not make them AGI. Be cautious with any answer option that claims otherwise.
• Weak and strong AI (Searle) differ from narrow and general AI. One is about whether machines can have minds, the other is about scope of ability.
• More autonomy is not always better. Questions about risk and governance usually favour human oversight for high-stakes decisions.

6. Why this matters for your AWS exam

The AI Practitioner exam expects you to explain basic AI concepts and terminology (Amazon Web Services, n.d.), so you should be able to define narrow AI and AGI and recognise that everyday AI applications are narrow. Check the current exam guide for exact wording and weightings. The vocabulary of agents matters more for the Developer exam, where designing and securing agent-based applications is part of the work, and this post gives you the foundation for those later posts.

7. Quick check

1. Which type of AI describes every system deployed today, in the technical sense?
2. What is the main difference between narrow AI and artificial general intelligence?
3. Is agentic AI a step on the way to superintelligence? Explain.
4. Name the two ways of describing AI that this post asked you to keep separate.
5. Give one reason a fixed workflow might be chosen over an agent.

Answers: (1) Narrow AI. (2) Narrow AI is built for specific tasks, while AGI would match human flexibility across many tasks and transfer learning to new situations. (3) No. Agentic AI describes how a system is used (pursuing goals with some autonomy), not how general its intelligence is, and agents are built from narrow components. (4) How capable or general the system is, and how it is used. (5) Any one of: lower cost, lower latency, easier testing, more predictable behaviour, easier governance.

What’s next

In the next post we will look at common AI use cases across industries, and just as importantly, the situations where AI is the wrong tool.

References

Amazon Web Services. (n.d.). AWS Certified AI Practitioner (AIF-C01) exam guide. Retrieved [Month Day, Year], from [URL of the current official exam guide]

Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press.

Goertzel, B., & Pennachin, C. (Eds.). (2007). Artificial general intelligence. Springer.

Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391–444. https://doi.org/10.1007/s11023-007-9079-x

Russell, S., & Norvig, P. (2021). Artificial intelligence: A modern approach (4th ed.). Pearson.

Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–424. https://doi.org/10.1017/S0140525X00005756

Siadati, Saman. (2024). Artificial Intelligence: Concepts to Applications. 10.5281/zenodo.17008758.