An AI agent is an AI model that can take actions to reach a goal, instead of only answering questions. You tell it what you want done. It plans the steps, uses tools such as a web browser, your files or a terminal, checks each result, and keeps going until the job is finished or it needs you.
That is the plain answer to what an AI agent is: a model, plus tools, plus a loop, plus a goal. This guide explains each part, follows one real task from start to finish, and shows how to try an agent without risking anything you care about.
- An AI agent is a model plus tools plus a loop plus a goal.
- It works in a cycle: look, plan, act, check, and repeat until the goal is met.
- You choose how much freedom it gets: suggest only, act with your approval, or act alone.
- Agents are strong at multi-step jobs with a clear finish line, and weak at vague goals and long unsupervised runs.
- Start with a low-stakes task, a supervised mode and nothing you cannot undo.
What is an AI agent made of?
Think of an agent as a new assistant. It needs a brain, access to your tools, a habit of checking its work, and a clear brief. Each maps to one part:
| Part | What it does | Example |
|---|---|---|
| Model | Reads, reasons and decides the next step | A large language model, like the ones behind ChatGPT, Claude or Gemini |
| Tools | Lets it do things outside the chat | Web search, reading files, running code, clicking in a browser |
| Loop | Makes it keep going after each step | Try, look at the result, adjust, try again |
| Goal | Tells it when to stop | “Every page passes the dark mode check” |
The model gets the attention, but the other three turn it into an agent. A brilliant model with no tools can only talk. With tools but no loop, it takes one action and stops. Without a clear goal, the loop cannot tell when it is finished. What it remembers from one task to the next is a separate question, and our guide to AI agent memory covers how that works.
Tools are often plugged in through the Model Context Protocol, an open standard that lets one agent connect to many apps and data sources. Our guide to MCP explains how it works.
How an AI agent works: look, plan, act, check
Every agent runs the same basic cycle. Here it is on one concrete task: “Add a dark mode to my website.”
Look
The agent reads your request, then the project: the page templates, the stylesheet and how colors are set. It notices the colors are typed out by hand in dozens of places.
Plan
It writes a short plan: gather the colors into one place, add a dark set, put a toggle in the header, and remember each visitor’s choice. If you asked to see plans first, it stops here for your OK.
Act
It makes the changes one at a time: the shared colors, the dark palette, the toggle button.
Check
It starts the site, switches to dark mode and takes screenshots. The footer text is now dark gray on black and impossible to read. That is a failed check, so the loop goes around again.
Repeat until done
It fixes the footer, checks every page again and runs the tests. Everything passes, so it stops and reports what it changed, what it checked, and one question for you: should dark mode follow the system setting?
The check step is what separates an agent from a script that runs blindly. Anthropic’s guidance for agent builders makes the same point: at each step the agent needs “ground truth” from its environment, such as the result of running code, to judge its progress.
A well-built agent also knows when to stop. It stops when the goal is met, when it has tried enough times without progress, or when it reaches a decision that belongs to you.
How much freedom should an agent get?
Autonomy is a setting, not a fixed trait. Most agents let you choose how much they may do on their own, and the options boil down to three levels.
| Suggest | Act with approval | Act alone | |
|---|---|---|---|
| What it does | Proposes steps or changes | Takes steps, pauses before important ones | Works through the whole task within limits |
| Who presses the button | You, every time | The agent, except at the pauses | The agent |
| Speed | Slow | Medium | Fast |
| Risk | Lowest | Low, if the pauses are in the right places | Highest |
| Good for | Trying a new agent, sensitive work | Most everyday tasks | Repeat jobs you have already watched it do well |
For most people, “act with approval” is the right default. It is fast enough to be useful, and the pauses land where a mistake would hurt: sending, paying, deleting and publishing. Move a task to “act alone” only after the agent has done that exact kind of task well a few times.
If you are still deciding whether you need an agent at all, our comparison of an AI agent vs a chatbot walks through three everyday requests handled both ways.
What AI agents are good at today, and where they struggle
- Multi-step jobs with a clear finish line
- Work it can check itself: tests pass, totals match, the page renders
- Repetitive changes across many files, rows or pages
- Pulling many sources together into one report or document
- Coding tasks in a project with good tests
- Vague goals, such as “make the site better”
- Long unsupervised runs, where small errors pile up
- Taste and judgment calls that depend on context only you have
- Information it cannot reach, such as a login it was never given
- Web pages or files written to trick it, known as prompt injection
- Knowing when it is wrong: agents can report success confidently
The trend is steep. In March 2025, METR, a nonprofit that measures AI capabilities, found that the length of tasks agents could finish had been doubling roughly every seven months for six years. By its May 2026 report, the strongest agents it tested had essentially maxed out its test suite: only a handful of tasks longer than eight hours were still unsolved.
Two cautions keep that in perspective. METR’s tasks are mostly software, machine learning and security work, and its headline measure is the task length at which an agent succeeds half the time. Your messy real-world errand is not a clean test task, so give agents short, checkable steps.
Real AI agents you can use in 2026
As of September 2026, agents come in three broad kinds. Many products now blend them.
| Kind | Examples | What it does |
|---|---|---|
| Work agents | ChatGPT Work (OpenAI), Claude Cowork (Anthropic) | Pulls context from your files and apps, then produces documents, spreadsheets, slides and reports |
| Coding agents | Claude Code, OpenAI Codex, Cursor, Cyborb (ours) | Reads a codebase, writes and runs code, and fixes what fails |
| Computer use agents | Browser and desktop agents | Operates apps by looking at the screen and using the mouse and keyboard |
The lines blur quickly. Claude Cowork, for example, can work in your local files, connected apps and a built-in browser. The third kind is the most general, since it can use almost any app a person can. Our guide to computer use agents covers how they see and click. Complex jobs sometimes split across several agents at once; our guide to multi-agent systems covers when that split earns its cost.
How to try an AI agent safely
Your first agent task should be boring, reversible and easy to check. Save the ambitious jobs for when you know how the agent behaves. If you would rather build one yourself, our tutorial on building your first AI agent does it in about 100 lines of Python.
A clear brief does half the safety work. This one is safe to try on a folder of copied photos:
Goal: Rename the photos in Pictures/Trip-copy so each name starts with the date taken, like 2026-08-14_001.jpg. Done when: Every file is renamed and none are missing. Count the files before and after. Limits: Only touch files in Pictures/Trip-copy. Do not delete or move anything. Before you start: Show me your plan and five example renames, then wait for my OK.
For the full risk picture, including prompt injection and sandboxes, read is it safe to let an AI agent control your computer?
FAQ
What is an example of an AI agent?
A coding agent fixing a bug is a classic example: it reads the code, makes a change, runs the tests and tries again until they pass. Work agents such as ChatGPT Work and Claude Cowork, which turn your files into finished documents, are another.
What is the difference between an AI agent and a chatbot?
A chatbot writes a reply and waits for you. An agent takes actions toward a goal, checks the results and keeps going. Many assistants now do both, depending on the mode you choose.
What is agentic AI?
It is the umbrella term for AI that acts toward goals on its own, rather than only generating text when asked. When a product calls itself agentic, ask what it can access, what it can do, and when it asks you first.
Do you need to know how to code to use an AI agent?
No. You describe the goal in plain language. Coding knowledge helps you judge a coding agent’s work, but work agents and many desktop agents are built for people who never write code.
- An AI agent is a model plus tools plus a loop plus a goal.
- The check step, looking at real results after each action, is what makes it reliable.
- Choose the level of freedom per task, and default to acting with approval.
- Agents shine on multi-step, checkable work and stumble on vague goals and long unsupervised runs.
Read next: how to automate repetitive tasks with AI, with five worked examples and prompts.
- Building effective agents, Anthropic, December 2024
- Measuring AI ability to complete long software tasks, METR, March 2025
- Frontier risk report (February to March 2026), METR, May 2026
- Task-completion time horizons of frontier AI models, METR, May 2026
- OpenAI launches ChatGPT Work, deepening race for workplace AI tools, Reuters via BNN Bloomberg, July 2026
- Claude Cowork, Anthropic
- How we contain Claude across products, Anthropic, May 2026
