Introduction to AI agents

Who this is for and what you will be able to do

This guide is for readers who have used a large language model and want to know what an AI agent is, how it differs from an ordinary AI application, and when building one makes sense. By the end you should be able to explain what an agent is, know when to use one and when not to, and sketch a basic agent design for a real problem ^[1].

What an AI agent is

An AI agent is a system that lets a large language model (LLM) actually do things. It gives the model tools and knowledge to act on the world, instead of only responding to prompts ^[1].

That definition has three parts. First, an agent is a system, not a single model. Every agent has three core pieces:

• The environment is the space the agent works in. For a travel booking agent, it is the booking platform itself ^[1].
• Sensors are how the agent reads the current state of its environment. A travel agent might check hotel availability or flight prices ^[1].
• Actuators are how the agent acts. A travel agent might book a room, send a confirmation or cancel a reservation ^[1].

Second, the LLM is what makes modern agents powerful. Agents existed before LLMs, but LLMs understand natural language, reason about context, and can turn a vague request into a concrete plan of action ^[1].

Third, the agent can act. Without an agent system, an LLM only generates text. Inside an agent system, it can carry out steps such as searching a database, calling an API or sending a message ^[1].

What tools and memory an agent has

The tools an agent can use depend on two things: the environment it runs in, and what the developer chooses to give it. A travel agent might be able to search flights but not edit customer records. The wiring decides ^[1].

An agent can also have two kinds of memory. Short-term memory is the current conversation. Long-term memory can be a customer database or a record of past interactions. In the travel example, the agent might remember that a customer prefers window seats ^[1].

The seven types of AI agent

Agents are often grouped into seven types. The travel booking agent serves as the running example for each one ^[1].

1. Simple reflex agents follow hard-coded rules. They have no memory and no planning. Forwarding a complaint email to customer service is the example ^[1].
2. Model-based reflex agents keep an internal model of the world and update it as things change. They might track historical flight prices and flag routes that suddenly become expensive ^[1].
3. Goal-based agents have a goal in mind and work out how to reach it step by step. They might book flights, a car and a hotel to get the traveller from where they are to the destination ^[1].
4. Utility-based agents weigh tradeoffs to find the best solution, not just any solution. They can balance cost against convenience to find the trip that scores highest for you ^[1].
5. Learning agents get better over time by learning from feedback. For example, they can adjust future booking recommendations based on post-trip survey results ^[1].
6. Hierarchical agents have a high-level agent that splits work into subtasks and hands them to lower-level agents. A cancel-trip request might be split into cancelling the flight, the hotel and the car rental ^[1].
7. Multi-agent systems (MAS) contain several independent agents that work together or compete. Cooperatively, separate agents can handle hotels, flights and entertainment. Competitively, agents can compete to fill hotel rooms at the best price ^[1].

When to use an AI agent

Just because you can use an agent does not mean you always should ^[1]. Agents fit best in three situations ^[1]:

• Open-ended problems, where the steps cannot be written in advance and the LLM has to work out the path as it goes ^[1].
• Multi-step processes, where the task needs tools across several turns rather than one lookup or one generation ^[1].
• Improvement over time, where the system should get better from user feedback or signals from its environment ^[1].

When not to use agents is covered in a later lesson on building trustworthy AI agents ^[1].

Start by defining what the agent can do

When you build an agent, the first step is to define what it can do: its tools, actions and behaviours ^[1]. Only after that do prompts and orchestration come into play.

Two terms come up often. Agentic patterns are reusable strategies for prompting and orchestrating LLMs, so that multi-step work is more scalable and reliable. Agentic frameworks provide ready-made templates, tools and infrastructure. They make it easier to connect tools, observe what an agent is doing and debug it, and get several agents to collaborate ^[1].

Platform and framework used in the examples

The examples in this series use Microsoft Foundry Agent Service as the main platform. The platform supports models from providers such as OpenAI, Mistral and Meta (Llama), licensed data from providers such as Tripadvisor, and standardized OpenAPI 3.0 tool definitions ^[1]. For building the agents themselves, the series focuses on the Microsoft Agent Framework (MAF), which is aimed at production-ready agents ^[1].

Where the code samples are

The sample code consists of a Python notebook, code_samples/01-python-agent-framework.ipynb, and a .NET file, code_samples/01-dotnet-agent-framework.md. Both are built on the Agent Framework ^[1]. This article covers concepts only and has no install steps. Run the samples when you reach the later material that needs them.

What you should be able to do afterwards

After reading this, you should be able to explain what an agent is and how it differs from a regular AI solution, decide whether a task calls for an agent, and sketch a basic design for a real problem, such as a travel booking service with sensors, actuators and a clear goal ^[1].

Adapted from lesson 1 of Microsoft's open-source course *AI Agents for Beginners*^[1], released under the MIT License.