What Are AI Agents?

What are AI agents?

In the previous courses, you learned how to use AI tools, write prompts, and understand generative AI. Now we're entering the next level: AI agents. Systems that don't just answer but act, plan, and complete tasks on their own. Welcome to the paradigm shift that really got going in 2025.

But what exactly distinguishes an AI agent from a normal chatbot? And why is everyone suddenly talking about "agentic AI"? In this lesson, you'll get the basics you need to understand the whole topic in depth.

Did you know? The term "agent" comes from computer science and means "acting entity." It has been used in AI research since the 1980s. But only in 2025 did the concept become practical, thanks to better language models, tool use capabilities, and longer context windows. In October 2024, the market research company Gartner put "agentic AI" in first place among its top technology trends for 2025.

Chatbot vs. AI agent

To understand AI agents, a clear comparison helps. A classic chatbot, whether rule-based or LLM-powered, works on the principle of prompt in, answer out. You ask a question, and the model answers. Done. Each interaction stands on its own.

An AI agent goes fundamentally further: it receives a goal, breaks it down into substeps, uses tools (APIs, databases, web search, code execution), checks its results, and corrects itself. It acts autonomously, within defined limits.

Classic chatbot:

You ask: "What will the weather be like in Zurich tomorrow?" The chatbot answers based on its training knowledge (which may be outdated) or says: "I don't have access to current weather data."

AI agent:

You say: "Plan my outdoor day in Zurich tomorrow." The agent calls the weather API, checks your calendar availability, looks for suitable outdoor activities, takes the weather forecast into account, and creates a concrete day plan with time slots and alternative suggestions in case of rain.

The three core characteristics

Every real AI agent has three fundamental properties that distinguish it from a simple LLM chat:

1. Autonomy: The agent makes its own decisions about its next steps. You specify the goal, not the way to get there. It chooses for itself which tools to use, in what order to proceed, and when it is finished.

2. Goal orientation: Instead of reacting to individual prompts, the agent pursues an overarching goal. It keeps the context across several steps and works systematically toward the result.

3. Tool use: The agent can use external tools: call APIs, search the internet, execute code, read and write files, query databases. This extends its capabilities far beyond pure text knowledge.

Practical tip: If you want to assess whether a system is an AI agent in the sense of this course, ask these three questions: Can it plan several steps on its own? Can it use external tools? Can it check and correct its own results? If the answer to all three is yes, you are dealing with an agent.

The perception-reasoning-action loop

AI agents work in a continuous cycle called the perception-reasoning-action loop:

Perception: The agent takes in information: your assignment, the results of its previous actions, error messages, new data from APIs. It "sees" where it stands.

Reasoning: Based on what it has perceived, the agent considers: What is my goal? What have I achieved so far? What is the best next step? This is where the strength of the LLM comes into play: logical thinking, planning, prioritization.

Action: The agent carries out a concrete action: an API call, code execution, a database query, a web search. The result flows back into perception, and the loop starts over.

This cycle repeats until the goal is reached or the agent recognizes that it isn't making progress and needs human help.

Types of AI agents

Not all agents are the same. In research and practice, three main types are distinguished:

Reactive agents

React directly to inputs without long-term planning. They follow predefined rules: if X, then do Y. Fast and predictable, but not very flexible.

Example: A customer service bot that categorizes incoming emails by keywords and sends prepared answers.

Strengths: Fast, predictable, easy to debug.

Weaknesses: Cannot deal with unexpected situations.

Deliberative agents

Plan ahead, develop strategies, and adjust their plans based on new information. They use chain-of-thought reasoning to break down complex tasks.

Example: A research agent that creates a market analysis: it plans its research strategy, collects data from different sources, analyzes it, and creates a structured report.

Strengths: Can solve complex, multi-step tasks.

Weaknesses: Slower, higher resource consumption, harder to control.

Hybrid agents

Combine reactive and deliberative elements. For routine tasks, they react quickly according to rules; in complex situations, they switch to planning mode.

Example: Claude Code acts quickly and directly on simple code changes. For larger refactorings, it plans a multi-step strategy with tests and validation.

Strengths: Flexible, efficient, practical.

Weaknesses: More complex to develop and test.

Examples from practice

AI agents are not a thing of the future; they are in use today:

Devin (Cognition): An AI software developer that writes code, debugs, creates tests, and opens pull requests on its own. It works in its own development environment and can work autonomously on features for hours.
Claude Code (Anthropic): A coding agent that runs directly in the terminal, reads and writes files, executes Git commands, and can implement complex software projects.
AutoGPT / BabyAGI: Early open-source agents (2023) that showed what is possible, even though they were still unreliable. They laid the foundation for today's agent movement.
Browser agents: Agents such as OpenAI Operator (early 2025, later integrated into ChatGPT) control a web browser and complete online tasks, from orders to research.

Example: Imagine you ask an AI agent: "Analyze our last 100 customer reviews and create an improvement plan." The agent would (1) load the reviews from your database, (2) carry out a sentiment analysis, (3) identify recurring themes, (4) prioritize the themes by frequency and severity, (5) suggest concrete improvement measures, and (6) format the whole thing as a structured report. A chatbot would have said: "Please paste the reviews in here."

Why the breakthrough now?

The concept of AI agents has existed for decades. Why are they only working now? Three factors came together in 2025/2026:

Better base models: The current models from OpenAI, Anthropic, and Google are powerful enough to handle multi-step reasoning tasks reliably. Earlier models made too many mistakes in planning.

Tool use as standard: All major providers (OpenAI, Anthropic, Google) have integrated native tool use interfaces into their models. Agents can now reliably use external APIs, databases, and code execution environments. In addition, the Model Context Protocol (MCP) has become established as an open standard for connecting AI agents to external tools and data sources, comparable to a "USB-C for AI applications."

Longer context windows: From a few thousand tokens (2022/2023) to usually 128,000 to more than 1 million tokens today. Agents can retain more context, process larger files, and keep longer chains of tasks in "memory."

Caution: AI agents are powerful, but not infallible. They can get stuck in loops, make wrong decisions, or use tools incorrectly. Never give an agent uncontrolled access to critical systems (production databases, financial systems, sending emails to customers). Always set limits and implement safety mechanisms; more on this in later lessons.

What fundamentally distinguishes an AI agent from a classic chatbot?

Correct! The three core characteristics of an AI agent are autonomy (acting on its own), goal orientation (multi-step planning), and tool use (APIs, code, databases). This fundamentally distinguishes it from the prompt-and-answer pattern of a chatbot.

Not quite. Neither the size of the model nor the target group is the decisive difference. AI agents are defined by three core characteristics: autonomy (acting on their own), goal orientation (multi-step planning), and tool use (APIs, code, databases). A chatbot only reacts to individual prompts, while an agent pursues a goal over several steps.

Key Takeaways:

AI agents act autonomously, pursue goals, and use external tools, unlike chatbots, which only react to individual prompts.
The perception-reasoning-action loop is the core mechanism: perceive, reason, act, in a loop until the goal is reached.
There are three types of agents: reactive (rule-based), deliberative (planning), and hybrid (combined).
The breakthrough in 2025/2026 was made possible by better models, native tool use interfaces, and larger context windows.
AI agents are powerful but need safety limits: uncontrolled autonomy is risky.

This lesson is part of a course with final exam and certificate. View the course "AI Agents & Automation"