AI agent means an AI that carries out a task for you, step by step, instead of only answering a question. It can use tools such as a web browser, your email or your calendar, and it decides the next step itself until the job is done.
A chatbot is a clever friend on the phone: it tells you how to book a train, and you do the booking. An agent is an assistant you have handed your keys to: it opens the website, fills in the form and books the train for you.
In a sentence: “I set up an AI agent to sort my inbox overnight: it files the newsletters, drafts replies to the easy emails, and leaves anything about money for me to check.”
What the word means
The plain point: the difference between an agent and an ordinary chatbot is who decides the steps. With an agent, the AI does.
Anthropic, the company that makes Claude, draws this line in its own guide. It calls an agent a system where the AI directs its own process and its own use of tools.1
The same guide calls the simpler version a workflow: the AI follows steps a person wrote out in advance. Think of a recipe card (a workflow) versus a cook who decides what to make from what is in the fridge (an agent).
An LLM, used in that quote, is a large language model: the text-writing engine inside tools like ChatGPT, Claude and Gemini. See LLM meaning.
Where the word comes from
The word is much older than today’s chatbots. “Agent” was already an established idea in computer science by the mid-1990s: a 1995 survey by Michael Wooldridge and Nick Jennings opened by saying so.2 What is new is the engine inside: today’s agents are usually built around a large language model that decides the steps.1
A simple example
The plain point: you give an agent a goal, not a question.
- Chatbot: “Which flights go to Lisbon on Friday?” It writes a list. You do the rest.
- Agent: “Book me a morning flight to Lisbon on Friday, window seat.” It opens a travel site, compares flights, fills in your details and, if you allow it, pays.
The second request involves many steps the AI chooses for itself: which site to open, what to click, when it is finished.
Advice you can ignore. An action has already happened. That is the whole difference, and it is why agents need more care.
Why it matters to you
The plain point: an agent can save you time, but it acts with your accounts, so its mistakes are real.
- It can be tricked. Text hidden on a web page or in an email can try to give the agent new orders. This is called prompt injection, and there is no complete fix.
- It can be confidently wrong. The AI inside it can still make things up. See hallucination meaning.
- The real protection is what it can touch. An agent that cannot pay cannot be tricked into paying. Permission tiers for agents explains how to set this.
Often confused with
- Chatbot: answers in a chat window and stops. It does not act in other apps.
- Agentic AI: the wider idea of AI that works towards goals with little supervision. An agent is one product built that way. See agentic AI meaning.
- Automation: a fixed rule (“every Monday, send this report”). It never decides anything new. An agent does.
Questions people ask
What is the difference between an AI agent and a chatbot?
A chatbot answers and waits for you. An agent takes actions in other apps and decides its own next step until the job is done. Anthropic draws the line at who directs the process.1
Do I need an AI agent for this?
Often not. If a single question, or a fixed set of steps, gets the job done, that is simpler and easier to check. Anthropic gives builders the same advice: “we recommend finding the simplest solution possible, and only increasing complexity when needed.”3
Are AI agents safe to use?
They can be, with limits. Because an agent acts on its own over many steps, one early mistake can carry through to the end, so give it only the access the task needs and keep approval for anything you cannot undo.3
Learn more on this site
- The full explainer: AI agents for beginners covers what agents can do today and how to start safely.
- The products: Meta Muse vs Gemini Spark vs Grok Bot, compared.
- Agents on the web: Letting an AI browse for you.
- Writing instructions for one: Prompting agents and agent guardrails.
- Other words: the AI glossary and AI explained for complete beginners.
- Related: Types of AI, Explained.
- Related: What Is a World Model in AI? World Model Meaning in Plain English.
- Related: How to Prompt AI Agents.
- Related: Chatbot vs AI Agent vs Agentic Workflow.
- Related: What Is an Agentic Workflow? Agentic Workflow Meaning in Plain English.
- Related: What Is Deep Research in AI? Deep Research Meaning in Plain English.
- Related: What Is an Orchestrator in AI? Orchestrator Meaning in Plain English.
- Related: What Is a Subagent? Subagent Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is the main difference between an AI agent and an ordinary chatbot?
- The agent always writes shorter answers
- The agent only works on a phone
- The agent decides the steps itself and acts in other apps
Show the answer
Answer: The agent decides the steps itself and acts in other apps. A chatbot answers and waits; an agent takes a goal, chooses its own steps and acts in other apps until the job is done.
2. Why does a mistake by an agent matter more than a mistake by a chatbot?
- An agent’s mistake is an action already taken
- An agent’s mistakes are harder to read on screen
- An agent’s mistakes are always about money
Show the answer
Answer: An agent’s mistake is an action already taken. A chatbot’s mistake is advice you can ignore, but an agent’s mistake is an action that has already happened.
3. What does the page call the real protection when you use an agent?
- Asking it politely to be more careful
- Limiting what it can touch, so it cannot pay if the task does not need it
- Only using it late at night
Show the answer
Answer: Limiting what it can touch, so it cannot pay if the task does not need it. The real protection is what it can touch: an agent that cannot pay cannot be tricked into paying.
Related terms: AI model · Algorithm (in AI) · Artificial intelligence (AI) · Neural network · Prompt · Token
- 1Anthropic, Building effective agents, 19 Dec 2024, read at source 1 Oct 2026“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage”. ↑a“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage”. ↑b“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage”. ↑c
- 2Wooldridge and Jennings, Intelligent Agents: Theory and Practice, Knowledge Engineering Review 10(2), 1995, read at source 12 Oct 2026“The concept of an agent has become important in both Artificial Intelligence (AI) and mainstream computer science.” ↑
- 3Anthropic, Building effective agents, 19 Dec 2024, read at source 12 Oct 2026“we recommend finding the simplest solution possible, and only increasing complexity when needed.” ↑a“The autonomous nature of agents means higher costs, and the potential for compounding errors.” “We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.” ↑b
The comparisons on this page (friend on the phone, house keys, recipe card and cook) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK