AI agents explained: what they are, how they differ from chatbots, and where you'll meet one

AI agent is one of the most-used phrases in AI right now, and it means something more specific than “chatbot with extra steps”. This guide explains what makes something an agent rather than an assistant that just answers questions, where you’re already using one, and how to use them safely. It draws on how the leading vendors describe their own agent features.

Chatbot vs agent

The line isn’t always sharp. Most assistants today, including ChatGPT, Claude and Gemini, can act as either, depending on the mode or tool you choose.

Where you’re already using an AI agent

• Everyday tasks: Claude Cowork and ChatGPT Work can browse, fill in forms and work across apps for you, with approval modes that decide how much they do without asking first.
• Always-on personal agents: OpenAI’s dots and Google’s Gemini Spark keep working between conversations on their own cloud computers; we compare them with Claude in dots vs Gemini Spark vs Claude Cowork.
• Research and browsing: Perplexity Computer breaks a goal into tasks and works through them with its own browser and file system, and Chrome’s auto browse feature does something similar for Google AI Pro and Ultra subscribers; see our guide to the best AI browsers.
• Coding: Claude Code, OpenAI’s Codex and GitHub Copilot’s agent mode plan a change, edit several files, run tests and can open a pull request. Our guide to the best AI coding assistants compares them.
• The smart home: Alexa+ describes some of its capabilities as agentic, for example completing a task through a partner such as Thumbtack rather than just answering a question about it.
• Business software: tools such as Salesforce’s Agentforce and HubSpot’s Breeze agents apply the same idea to sales and support, resolving a customer conversation or qualifying a lead with limited human involvement, usually billed by usage rather than a flat seat price.

What makes an agent an agent

Across these products, three things tend to be true:

1. It plans. Given a goal, it breaks the goal into steps rather than answering in one shot.
2. It uses tools. A browser, a file system, a calendar, a connected app or a piece of business software, not just its own knowledge.
3. It acts with limited supervision. It runs several steps before reporting back, rather than waiting for you to approve every action.

Approval modes: how much an agent does without asking

Most consumer agents let you control how much runs unsupervised. Claude Code’s permission modes are a clear example: a default mode that reads but asks before editing or running commands, a mode that accepts edits automatically while still asking about commands, and a fully autonomous mode meant only for isolated environments. Our guide to how to use Claude Code explains each mode in detail, and the same trade-off, more autonomy against more risk, shows up in every agent product.

Staying safe with an AI agent

1. Review the plan before it runs, when the tool offers one.
2. Keep confirmation on for irreversible actions: purchases, sent messages, deleted files.
3. Treat pages and documents an agent reads as untrusted. A web page can contain hidden text aimed at an AI agent rather than at you, so watch what an agent does right after it reads an unfamiliar source.
4. Grant the narrowest access a task needs, and check what you’ve connected afterwards.
5. Keep sensitive accounts, such as banking, out of an agent’s reach unless the specific task requires it and you trust the tool.

For a broader look at picking an assistant in the first place, see our guide to choosing an AI assistant, and for the plans behind the leading agents, read Claude Pro vs Max, ChatGPT Pro explained and Google AI Plus vs Pro vs Ultra.