An AI agent is a system that is given a goal, decides which steps to take, carries them out with tools — such as clicking and typing in a browser — checks the result and continues until the goal is reached or it needs a human.
A simple definition
A chatbot produces text. An AI agent produces outcomes. You give it a goal (“book a table for two at eight”), and it works in a loop: look at the current situation, choose an action, perform it, look again. The loop ends when the goal is met, when the agent gets stuck, or when it reaches a point where a person must decide.
Three ingredients make something an agent rather than a chat window: a goal, tools it can use to change things, and the autonomy to choose its next step.
How an agent works, step by step
- Observe. The agent receives a description of the world — for a browser agent, the page’s buttons, links, fields and text.
- Decide. A language model picks the next action that moves toward the goal.
- Act. The action runs: click, type, scroll, open a page.
- Check. The agent observes the new state and repeats.
Good agents add a fifth step: ask. Before something irreversible, they stop and let a human approve. This is called human in the loop.
Kinds of AI agents
- Browser agents operate websites for you. They come as extensions (Hark, Claude in Chrome), as features of a browser (Gemini in Chrome, Copilot in Edge) or as whole AI browsers (Comet, Opera Neon).
- Coding agents read and change software projects.
- Computer-use agents control a whole desktop.
- Workflow agents connect business apps through APIs.
This guide is about the first kind, because it is the one most people can use today without any setup.
What a browser agent can do for you
Think of errands that are easy to describe and tedious to do: booking a table, filling a long form, finding the cheapest item on a page, unsubscribing from a newsletter. An agent handles the clicking and typing; you supply the intent and the final yes.
Limits you should know
Agents make mistakes. They can misread a page, choose the wrong option or be misled by text planted on a website (prompt injection). They are slower than an expert human on a familiar site and cannot do what requires you: logging in, passing a CAPTCHA, entering a card. The practical rule: use agents for low-stakes, well-defined tasks, and keep approval for anything that costs money or sends a message. More in Are AI browser agents safe?
Where Hark fits
Hark is a browser agent delivered as a Chrome extension. It has three modes — Chat, Agent and Tracker — acts only in the tab where you start a task, and asks for approval before any step that sends, submits, books, pays or deletes. It is a gentle way to try an agent: no new browser, no account, a free plan.