What it means
An AI agent is an AI system that works towards a goal by choosing steps, using tools and acting on the results, rather than only replying to messages.
A chatbot answers, and you decide what happens next. An agent is handed a task and takes the next steps itself. It might search the web, look something up in company files using RAG, run code or send an email. Then it checks what happened and picks its next move.
Most agents are built around a large language model. The model decides, and the tools act. A loop ties them together: the model chooses an action, software carries it out, the result goes back, and the cycle repeats until the task is done or something breaks.
Think of the difference between asking a friend for directions and asking them to drive you there. Directions are advice. Driving means they make the turns and pick a new route when a road is shut. That’s more useful, but you’re trusting their judgement for the whole trip. How much control you hand over is what people mean by levels of autonomy.
| Type | What it does | Who checks the work |
|---|---|---|
| Chatbot | Answers questions in conversation | You, before acting on the answer |
| Assistant with tools | Uses a tool when asked, such as a web search | You, after each request |
| Supervised agent | Carries out several steps, pausing before risky ones | You, at checkpoints |
| Autonomous agent | Works through a whole task alone | Often nobody, until it’s finished |
Where it came from
“Agent” is an old word in AI research. The usual account is that textbooks from the 1990s onwards framed AI as the study of agents: anything that senses its surroundings and acts on them. By that definition, even a thermostat is a very simple agent.
The modern sense arrived with large language models. Developers began connecting chatbots to tools and letting them run in loops. In 2023, open-source projects such as AutoGPT drew wide attention by letting a model work through its own task lists. They were exciting to watch and unreliable in practice, often going round in circles.
As models got better at using tools, “agent” became the industry’s favourite product word, and from 2024 it was all over launch announcements. The adjective “agentic” came with it. It already existed in psychology and sociology, meaning having agency. Tech borrowed it to mean “working like an agent”.
The word also got stretched. Some products sold as agents are ordinary chatbots or fixed automations with a new label. Critics call this “agent washing”, after greenwashing. Real agents can be useful, especially in software, where coding agents power a lot of vibe coding. The label alone doesn’t tell you much.
How people actually use it
- “Meet your AI agent: it books meetings, chases invoices and updates your CRM while you sleep.” Product page. The marketing pitch: a digital employee.
- “We’re piloting an agent for first-line support, but refunds still go to a human.” Work chat. A supervised agent with firm limits.
- “Is it actually agentic, or is it a chatbot with a button?” Social post. Scepticism about the label.
- “Tech firms race to build AI agents that can use your computer for you” News headline. The general-audience framing.
- “I asked the coding agent to fix the failing tests and it deleted them.” Developer joke with a real point: agents can meet the letter of a task and miss its intent.
In a sentence
Nina: The vendor says their product is an AI agent.
Omar: Ask what it does without a human clicking a button. If the answer is nothing, it’s a chatbot.
Jess: Did the agent finish the report?
Kai: It drafted it, pulled the sales figures and asked before emailing the client. Exactly right.
Ben: Everything’s “agentic” now.
Lou: My toaster is agentic. It decides when the toast is done.
Common misconceptions
- “An AI agent is a robot.” Most agents are just software. They act through apps, browsers and code, not arms and legs.
- “Agents have their own goals.” In practice, the goal comes from whoever set the task. The model picks each step to serve that goal.
- “More autonomy is always better.” Every unchecked step lets a mistake snowball. An agent can hallucinate a fact early and then act on it repeatedly. Often, an agent that asks first is the better design.
Related terms
- Agentic: the adjective, for AI that plans and acts rather than only responds.
- Tool use: a model calling outside functions such as search or email, also called function calling.
- Agent loop: the repeating cycle of choosing an action, doing it and reading the result.
- Human in the loop: a design where a person approves key steps before they happen.
- Multi-agent system: several agents with different roles, passing work between them.
Questions people ask
What is the difference between a chatbot and an AI agent?
A chatbot replies to messages and leaves any action to you. An AI agent is given a goal and takes steps towards it itself, such as searching, editing files or sending messages, then chooses its next move from the results.
What does agentic mean?
Agentic describes AI that acts with some independence, planning steps and using tools to finish a task. An “agentic workflow” is one where AI carries out several steps rather than producing a single answer.
Are AI agents safe to use?
An AI agent is only as safe as the permissions it’s given and the checks around it. An agent that reads your calendar is low risk, but one that can spend money or send emails needs approval steps and limits.
Why does every AI product call itself an agent now?
“Agent” became a strong selling word from 2024, so many companies apply it to products that are really chatbots or simple automations. The useful question is what the product can do on its own, and who checks its work.
The short version
An AI agent is AI that doesn’t stop at answering. It takes a goal, picks steps, uses tools and acts on the results, with more or less supervision. The idea is old and useful, but marketing stretches the label, so ask what a so-called agent can really do by itself.