💡 Think of it like this:
Imagine you ask a regular AI chatbot to "plan a holiday to Japan." It'll give you a nice list of suggestions — but that's it. You still have to book the flights, find the hotels, check your calendar, and coordinate everything yourself.
Now imagine you gave that same request to a personal assistant who could actually do things. They check your calendar, search for flights within your budget, book the hotel, create a day-by-day itinerary, and send you the confirmation — all on their own, only checking in with you when a decision is needed.
That's the difference between a regular AI model and an AI agent. A chatbot responds. An agent reasons, plans, and acts.
Key insight: The word "agentic" means the AI has agency — the ability to independently take actions in the real world, not just generate text. This is the next evolution beyond generative AI.
AI agents are advanced AI systems designed to autonomously reason, plan, and execute complex tasks based on high-level goals. Unlike traditional generative AI models that follow a basic "request-and-respond" pattern, AI agents go beyond by orchestrating resources, collaborating with other agents, and utilising tools to achieve multi-step objectives.
They're often referred to as "agentic AI" or "LLM agents" because they typically use a large language model as their "brain" — but they pair it with memory, planning, and external tools to actually do things.
<aside> 💡
You already know what LLMs and generative AI are from your Fundamentals notes. AI agents are what happens when you give those models the ability to take action — not just generate content.
The NVIDIA article mentions several tools that AI agents rely on. Here's what each one actually means:
Large Language Models (LLMs)
You already know these from your Generative AI notes. LLMs like GPT are the "brain" of the agent — they understand your request and coordinate the plan. But on their own, LLMs can only generate text from what they were trained on. The tools below extend their capabilities beyond that.
Retrieval Augmented Generation (RAG)
LLMs have a knowledge cutoff — they only know what was in their training data. RAG solves this by letting the agent retrieve real, up-to-date information from external sources (documents, databases, websites) before generating a response. The LLM then uses that retrieved information to craft an answer that's grounded in actual facts. Without RAG, agents can "hallucinate" — confidently generating answers that sound right but are factually wrong. RAG anchors the agent's responses in real data.
➡️ This is a big topic — see your full dedicated page: Retrieval-Augmented Generation (RAG) in your Application Domains section.
Vector Databases
A specialised database that stores data as vectors (lists of numbers that represent meaning) rather than traditional rows and columns. This lets AI systems search by conceptual similarity instead of exact keyword matching. When an agent uses RAG, it searches a vector database to find the most relevant information — your question gets converted into a vector, the database finds the closest matching vectors, and that information gets fed to the LLM.
➡️ See your full dedicated page: Vector Databases in your Application Domains section for the complete breakdown.
APIs (Application Programming Interfaces)
APIs are how software systems talk to each other. For an AI agent, APIs are the "hands" that let it interact with the outside world — pulling data from a sales database, sending an email, booking a flight, or triggering an action in another system. Without APIs, the agent could only think and talk but never actually do anything.
Frameworks
Agent frameworks (like LangChain, Llama Stack, or custom Python code) are the development platforms used to build agents. They handle the plumbing — how the LLM connects to memory, tools, and APIs — so developers can focus on what the agent should actually do rather than wiring everything together from scratch.
</aside>
An AI agent is built from four key parts working together:
What are Large Language Models? | NVIDIA Glossary
The large language model is the central coordinator. It interprets your request, reasons through what needs to happen, decides which tools to use, and plans the sequence of actions. Think of it as the project manager — it doesn't do all the work itself, but it decides who does what and in what order.
Agents need to remember things to work effectively:
Planning is what separates agents from simple chatbots. The agent breaks complex tasks into actionable steps: