Imagine two pizza places.
At Pizza A, you call and reach a person whose job is to recite the menu. You say "I want a large pepperoni." They say "We have small, medium, large, in pepperoni, mushroom, or cheese." You say "okay, large pepperoni." They say "We have small, medium, large…" That's it. They only talk. They cannot take the order.
At Pizza B, you call and reach a manager. You say "I want a large pepperoni." The manager pulls up the system, places the order, calls the dough station, checks the oven timer, dispatches a driver. Twenty minutes later they call you back: "Driver's stuck in traffic, we're giving you a free dessert, ETA fifteen more minutes." They talk and they act in the world.
Pizza A is a chatbot. Pizza B is an agent.
The idea, in plain English
A language model (often LLM — Large Language Model) is the AI "brain" that reads and writes text, like ChatGPT or Copilot. On its own it can only produce words — it can't click, read a file, or act. It's the engine inside both chatbots and agents; the difference is what you bolt onto it.
A regular language model (the "chatbot") only generates text. You type, it answers, conversation ends. It cannot click a button. It cannot read a file. It cannot make a decision and try again if it fails. It just produces words.
An agent is a language model wrapped in a loop that lets it do things. The agent has a goal, a set of tools it can call (search the web, send an email, run code, query a database), and the ability to look at what just happened and decide what to do next. The conversation isn't the product — the outcome is.
An AI agent is a system that uses a language model to autonomously decide and execute actions toward a goal, by invoking tools and reasoning over the results across multiple turns.
The three things that make something an agent
Whenever you're unsure if something is an agent or just a fancy chatbot, run this checklist:
- Goals. It is trying to achieve a specific outcome (book the flight, fix the bug, draft the email), not just produce text.
- Tools. It can do things beyond generating words — call APIs, read files, run code, control a browser, send messages.
- Decisions. It looks at what happened after each action and decides what to do next. If a tool fails, it tries something else.
Miss any one of those three, and you have a chatbot, not an agent.
A concrete example
Say you ask an AI assistant: "Find me a flight from Toronto to Lisbon next Friday under $700."
A chatbot might respond: "You can check Google Flights or Skyscanner for that route."
An agent would: (1) open a flight search tool, (2) query Toronto→Lisbon for next Friday, (3) filter under $700, (4) notice no direct flights match, (5) decide to expand the search to include one-stop options, (6) return three concrete options with prices and links. Same input. Very different output. Because the agent has goals + tools + decisions.
The cert-language version
You'll be tested on the formal definition, so memorize this exact phrasing:
In plain terms: an AI agent is software that runs on a language model and decides for itself what to do next to reach a goal — picking and using tools, reading the results, and looping over several turns until the goal is met. The "deciding and acting on its own, in a loop" part is what separates it from a chatbot.
Our definition · grounded in MS Learn — GH-600 exam page · fetched 2026-05-27
Three words to remember: goal, tools, turns. If a system has all three, it's an agent. If it only generates text, it's a chatbot.
Common confusions (read these or get them wrong on the exam)
- "ChatGPT is an agent." Not by itself. The base chat model is a chatbot. ChatGPT becomes an agent only when wrapped with tools (browsing, code interpreter, plugins).
- "An agent needs to be smart." No. Intelligence isn't the test. A simple goal-tool-decision loop using a small model is still an agent.
- "Agents are always multi-turn." Single-turn agents exist (one tool call, one result, done). The "multiple turns" in the definition means it can take multiple turns, not that it always does.
- "Workflow = agent." No. A workflow is a hand-written script that calls tools in a fixed order. An agent decides the order itself.
Ticks this lesson done on the home roadmap. Saved in this browser.