AI Agents: What They Are, Types and the Most Used (2026)
An AI agent is a program that pursues a goal by deciding its own steps and acting on them: running code, browsing, calling tools. A chatbot answers; an agent does the work and reports back. This guide covers the kinds of agents in use today and the open protocols that connect them to tools.
Updated Nov 12, 2026
What is an AI agent?
An AI agent is an artificial intelligence program that can make decisions, pursue goals and take actions with little human input. In practice, today's agents combine a language model with tools and a loop: the model plans a step, uses a tool, looks at the result and decides what to do next, until the task is done or it needs to ask you something.
AI agent vs chatbot
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers messages | Carries out tasks |
| Steps | One reply at a time | Many steps, chosen by itself |
| Tools | Usually none | Terminal, browser, files, MCP servers |
| Result | Text | Changes made: code edited, form filled, report written |
Types of AI agents
- Coding agents work in your code: Claude Code, Codex, Gemini CLI. See our
- Browser and computer-use agents operate websites and apps like a person
would, such as ChatGPT agent.
- Research agents search, read and write up a report from many sources.
- Business and support agents answer customers and handle routine work,
such as Salesforce Agentforce.
- Personal, self-hosted agents run on your own machine, such as OpenClaw.
AI agents people use in 2026
| Agent | Maker | Kind | Note |
|---|---|---|---|
| Claude Code | Anthropic | Coding (terminal) | Default Fable model is Claude Fable 5.2; supports MCP |
| Codex | OpenAI | Coding | Open-source CLI (Apache-2.0) and cloud agent; supports MCP |
| Gemini CLI | Coding (terminal) | Open source (Apache-2.0); supports MCP | |
| ChatGPT agent | OpenAI | Computer use | Released July 2025, works on a virtual computer |
| Devin | Cognition | Autonomous software engineer | First shown in March 2024 |
| Manus | Butterfly Effect | General-purpose | Independent company since August 2026 |
| OpenClaw | OpenClaw Foundation | Personal, self-hosted | Open source (MIT) |
| Agentforce | Salesforce | Business and support | Agents for customer service and sales |
How agents use tools: MCP
The Model Context Protocol (MCP) is the open standard most agents use to reach tools and data. An MCP server describes its tools, and the agent lists them and calls the one it needs. Because it is a standard, one server works in Claude Code, Codex, Gemini CLI, ChatGPT and many more. Browse the servers in our MCP directory, or start with what MCP is.
MCP vs A2A vs WebMCP
| MCP | A2A | WebMCP | |
|---|---|---|---|
| Connects | An agent to tools and data | Agents to other agents | A web page to the agent visiting it |
| Started by | Anthropic, November 2024 | Google, April 2025 | Microsoft and Google engineers, August 2025 |
| Governed by | Agentic AI Foundation (Linux Foundation) | Linux Foundation | W3C community group draft |
| Status | Standard in use | Version 1.0 released (March 2026) | Experimental, in Chrome trials |
More detail in MCP vs WebMCP and in our WebMCP directory.
Risks: prompt injection and permissions
An agent reads web pages, documents and tool results, and any of them can hide instructions meant to trick it. That is called prompt injection. Keep agents safe by giving them only the access a task needs, approving sensitive actions yourself and installing tools, servers and skills only from publishers you trust. To teach an agent a procedure, see the Agent Skills directory.
Frequently asked questions
What is an AI agent in simple terms?
Software that takes a goal and carries out the steps itself, using tools such as a browser, a terminal or an MCP server, instead of only replying with text.
What is the difference between an AI agent and a chatbot?
A chatbot replies to messages. An agent takes actions and works through multi-step tasks, checking its own results along the way.
What are the types of AI agents?
By use: coding, browser and computer-use, research, and business or support agents. Textbooks also classify them as reflex, model-based, goal-based, utility-based and learning agents.
What are examples of AI agents in 2026?
Claude Code, Codex, Gemini CLI, ChatGPT agent, Devin, Manus, OpenClaw and Salesforce Agentforce, among many others.
Is ChatGPT an AI agent?
ChatGPT is a chatbot, but its agent mode, released in July 2025, carries out multi-step tasks on a virtual computer.
MCP vs A2A: which do I need?
MCP connects an agent to tools and data; A2A lets agents talk to each other. They complement each other, and many systems use both.
Do AI agents need MCP?
No, but MCP is the common standard for tools, so one MCP server works across many different agents.
Are AI agents safe?
They can be tricked by prompt injection hidden in web pages or documents. Give them only the permissions they need and use tools from publishers you trust.