Start here · Beginner guide

What is an AI agent?

An AI agent is software that can take actions for you, not just answer questions. You tell it a goal, such as "book me a haircut tomorrow afternoon" or "fix this bug", and it works through the steps using the tools you allow it to use. This page explains the idea in plain English, and helps you decide whether you need one and where to begin.

Agent or chatbot: what is the difference?

A chatbot answers. You ask a question and it replies with text. You still do the work: you copy the reply, send the email, make the booking.

An agent acts. You give it a goal, it decides on steps, uses connected tools such as your email, calendar, files, a web browser or a code editor, and reports back. Some agents ask you before each action. Others carry on by themselves for as long as the job takes.

The line is blurry. Many products mix both. The useful question is not "is it an agent?" but "what can it do without me, and what does it need my approval for?"

How an agent works, in four steps

You set a goal. You type or say what you want, often in a messaging app or a chat window.
It plans. An AI model breaks the goal into steps and picks tools to use.
It acts. It calls those tools: reads an email, searches the web, edits a file, fills in a form, makes a phone call. What it can reach depends on the access you grant.
It checks and reports. It reviews what happened, asks you about anything it cannot decide, and shows you the result.

Two settings matter most. Access is which accounts and tools it can touch. Approval is which actions it must ask about first. For example, Anthropic documents that terminal Claude Code asks permission before changing files or running commands, and Comma's site says its computers stay read-only until you allow more access. Those are vendor statements, which we have not tested. See each agent's record for its sources.

What people use agents for

Vendors document agents for a handful of jobs. Each links to agents whose own pages describe it:

Writing and fixing software, such as Claude Code, Devin and Replit Agent.
Personal email, calendar and reminders, through assistants you message, such as Poke, Lucas and Instinct.
Errands and long-running tasks, such as Fo by Wajo, which describes booking and calling on your behalf.
Customer service, such as Fin.
Repeatable team workflows, such as ChatGPT workspace agents.
Building your own, for developers, for example with LangGraph.

Why you might want one

Time. Tasks that are repetitive, have clear steps and are easy to check are the best candidates: scheduling, chasing replies, drafting, routine code changes.
Follow-through. An agent can keep working after you close the chat, if the product is built that way.
Access to many tools in one place. One request can cross email, calendar and documents.

You may not need one. If you only want answers or a first draft, a chatbot is enough. If the task is high-stakes, hard to check, or involves money you cannot afford to lose, wait until you trust the controls.

What can go wrong

Mistakes made at speed. Vendors warn about this. Instinct's terms say safeguards may not prevent unintended actions, and Meta warns of inaccurate responses or unexpected actions from Muse.
Too much access. An agent that can read all your email can also be fed bad instructions hidden in an email. Start with the least access that does the job.
Data that stays. Disconnecting an app does not always delete what was already kept. Instinct's privacy policy and OpenAI's dots FAQ both say so.
Costs that grow. Many agents charge by usage, credit or outcome. A free plan is not always a free task. Our directory has a pricing filter, there are guides to free AI agents , building versus buying and Manus alternatives by job, and unknown prices are labelled as unknown.

Our privacy and approval checklist lists eight questions to ask before connecting anything.

Where to start

Write down one small, boring task you want off your plate.
Find agents documented for that job on the use cases page.
Filter the directory by tool type and pricing.
Read each record's limits, permissions and pricing notes, then compare two.
Connect one low-risk account, set approvals to "ask first" where offered, and review everything it did after the first week.

Words you will see

Agent: Software that pursues a goal by taking actions with tools.
Model: The AI system, such as a large language model, that does the reasoning. Agents are built around one or more models.
Tool or integration: A connection that lets the agent do something: read email, search the web, run code.
Permissions or access: What the agent is allowed to see and change.
Human in the loop: A person approves or reviews steps before they take effect.
Memory: Information the agent keeps between conversations. Ask what it remembers and whether you can edit or delete it.
Credits or usage-based pricing: You pay for how much the agent does, not a flat fee.
Framework: A toolkit for developers who build their own agents, as opposed to a ready-made product.

Product statements on this site come from the vendors' own pages, checked 1 October 2026, and are not independent tests. See the methodology.