What Is an AI Agent? A Practical Explanation With Real Examples

The short answer

An AI agent is software that uses a language model to work toward a goal by taking actions, looking at what happened and deciding what to do next, over and over, until the job is done or it needs you. A chatbot answers a question. An agent is given an outcome ("get the failing tests passing", "answer this customer and issue the refund if the order qualifies") and does the steps in between.

Three ingredients make the difference:

1. A model that can reason about a task and decide on a next step.
2. Tools the model can call: read and edit files, run a command, search the web, click in a browser, query a database, call an API.
3. A loop that feeds the result of each action back to the model so it can choose the next one.

Take the loop away and you have a chatbot. Take the tools away and you have a model that can only talk about the work.

What the loop looks like in practice

Suppose you ask a coding agent to fix a failing test. A typical run goes like this:

1. It runs the test suite and reads the failure.
2. It searches the codebase for the function named in the stack trace and opens the file.
3. It forms a theory, edits the code, and runs the tests again.
4. One test still fails. It reads the new error, adjusts the fix, and runs them a third time.
5. Everything passes. It shows you the diff and a summary of what it changed and why.

Nobody scripted those five steps. The agent chose each one from the result of the last. That is also why agents fail differently from ordinary software: a wrong theory in step 3 can send the following steps in the wrong direction, which is why good agents show their work and ask before doing anything hard to undo.

Agent, chatbot, model, workflow: the difference

The lines blur in real products. Many chat assistants now run an agent loop when you ask for something that needs tools, and many reliable "agents" in business are mostly fixed workflows with an agent step where judgement is needed.

A longer comparison, including agentic AI, skills and bots: AI agent vs chatbot vs agentic AI.

The main kinds of agent in use today

Coding agents read a codebase, edit files, run commands and open pull requests. This is the most mature category, because code gives an agent fast, objective feedback: tests pass or they do not. By package downloads the most used right now are OpenAI Codex (25.6M a week), Claude Code (14.8M a week), Pi (5.2M a week). See the full coding agent ranking.

General-purpose agents take an open-ended task such as research, a report or a booking and carry it out in a sandbox with a browser, a file system and code execution. Examples are Manus and the open-source personal agent OpenClaw. See general-purpose agents.

Browser and computer-use agents operate a browser or desktop the way a person does, by reading the screen, clicking and typing. They are useful for sites with no API, and are harder to make reliable than agents that work through APIs, because every page layout, pop-up and login is a chance to go wrong. See browser agents.

Customer support agents answer customers from a company's knowledge base and systems, resolve what they can and hand the rest to people. Many are priced by the conversation they resolve instead of by the seat. See support agents.

Frameworks and builders are what developers and operations teams use to make their own agents: code libraries such as LangGraph, and visual platforms such as n8n. See frameworks and automation platforms.

Voice agents answer and place phone calls: booking appointments, qualifying leads, handling first-line support. Hosted platforms such as Vapi and Retell AI bundle telephony and speech; open-source frameworks such as LiveKit Agents let developers assemble their own. See voice agents.

Sales agents research prospects, write outreach and book meetings, work that sales development reps used to do by hand. See sales agents.

AI agent examples

What agents are actually used for, with an example product for each. Each link goes to a page with live usage data.

For the leaders in each of these areas today, see the best AI agents by category.

What agents are good at, and where they break

Agents do well when a task has three properties: the steps can be done with the tools available, success can be checked (tests, a form that submits, a number that reconciles), and mistakes are cheap to undo. That describes a lot of software work, data clean-up, research and first-line support.

They struggle when success is a matter of taste, when the task depends on knowledge that is not written down anywhere the agent can read, or when one wrong action is costly. Long tasks compound small errors, so reliability drops as the number of steps grows. Agents that read untrusted content, such as web pages or incoming email, can also be steered by instructions hidden in that content, a problem known as prompt injection.

The practical consequence: give an agent a narrow goal, a way to check its own work, and limits on what it may do without asking.

How to judge an agent

Marketing pages all say the same things, so look at evidence instead:

• Adoption. Are people using it? Package downloads, extension installs and repository stars are public for many agents and are what this site tracks.
• Momentum and maintenance. Is usage rising or falling, and when was the last release? In the past year several well-known agents were discontinued, merged into other products or archived.
• Independent quality data. For coding agents there are public benchmarks and marketplace ratings. For most other categories there is little beyond vendor claims, so run a trial on your own tasks.
• Cost model. Flat subscription, per-token billing through your own API key, credits, or per-outcome pricing. The cheapest entry price is rarely the cheapest at real volume.
• Control. What can it do without asking, can you see each step, and can you run it on your own infrastructure?

Getting started

If you want to make your own rather than use a ready-made one, see how to build an AI agent. Otherwise, pick one small, checkable task you do often. Try the two leading agents in the relevant category on that same task, watch each step the first few times, and only then widen what you allow the agent to do on its own. The best AI agents by category and the full rankings are a good place to draw up that shortlist, and the glossary explains the vocabulary you will meet along the way.