The difference in one example
Ask a chatbot "help me invoice last week’s clients" and it will explain the steps or draft invoice text for you to paste. You still open the spreadsheet, the accounting app and your email.
Ask an agent the same thing and it opens the spreadsheet, finds the unbilled rows, fills each invoice in your accounting app, exports the PDFs and brings them back for your OK. Same model underneath, perhaps, but with tools and the freedom to choose its next step.
Side by side
| Chatbot | AI agent | |
|---|---|---|
| Output | An answer: text, code, an image | A finished task and a report |
| Who does the steps | You | The agent, in a loop |
| Tools | Few or none | Files, apps, browser, screen, accounts |
| Length of a turn | Seconds | Seconds to many minutes |
| Cost per request | One model call | Many model calls |
| Main risk | A wrong answer you might act on | A wrong action taken for you |
| Best guardrail | Check the answer | Limit access, approve important actions |
Why the line is blurring
Chat assistants have been gaining tools for years: web search, file analysis, running code, connecting to email. Each tool makes them a little more agent-like. Anthropic describes agents as models using tools based on feedback in a loop, and the key word is loop: the more a system decides and acts on its own across many steps, the more it is an agent. A chatbot that searches once and answers sits near the chat end; a system that works through a task for twenty minutes sits at the agent end.
When should you use a chatbot?
- You want to understand something: an explanation, a second opinion, a summary of a document.
- You want a draft to edit yourself: an email, a paragraph, a formula.
- The task is one step, or you want to keep every step in your hands.
