What is an AI agent, and how is it different from a chatbot?

An AI agent is a system in which a language model directs its own steps and its use of tools (a search, a website, a file), typically in a loop (Anthropic’s definition, 19 December 2024). If an AI writes you a list it is a chatbot or an assistant; if it prepares the shop by itself in several steps and stops to ask for your go-ahead, it is an agent.

Updated October 2026. Sources: Anthropic (Building effective agents, tool use, Claude in Chrome); Google (Google Cloud, Gemini API); NIST; OpenAI (guide to agents, function calling, cloud browser); MCP specification (principles, tools), all viewed on 20 October 2026.

What is an AI agent?

The sources do not use the same words (pages viewed on 20 October 2026, in alphabetical order):

• Anthropic. Separates workflows, where the code fixes the steps, from agents, where the model chooses them (19 December 2024).
• Google Cloud. Software that uses AI to pursue goals on behalf of its users, combining reasoning, planning and memory with some autonomy (page updated in 2026).
• NIST (United States). Systems that can plan and act on their own, with effects on real-world systems or environments (12 January 2026).
• OpenAI. Systems that carry out tasks for you independently; simple chatbots are not agents (undated guide).

Not every task needs an agent: Anthropic advises starting with the simplest solution.

Chatbot, assistant and agent: what is the difference?

How an agent works: goal, tools, steps

An agent works in a loop: goal, tool, result, next step. A hypothetical example with the weekly shop:

1. Goal. You say what you want, not how: “prepare the shop for two people”. It starts from a command from you or a discussion with you (Anthropic).
2. Plan. It breaks the task down: a menu, what you have at home, the list.
3. Tools. It picks a tool for each step. The model does not run anything itself: it names the tool and the data, and the software around it runs it (Anthropic, Google, OpenAI).
4. Check. It reads the result, corrects and tries again; it stops when the task is done or at a limit, such as a maximum number of loops (Anthropic).
5. Memory. It keeps the thread of the task and, where designed to, remembers preferences such as “no fish”; Google Cloud distinguishes short-term from long-term memory.
6. Your go-ahead. Before ordering or paying it should stop: OpenAI lists payments among the actions a person should review, and the MCP specification recommends that a person should always be able to deny the use of a tool.

What can go wrong?

Three risks, each described by the sources themselves (October 2026):

• Errors that add up. Anthropic writes that autonomy brings higher costs and the risk of errors piling up, and recommends testing in sandboxed environments; OpenAI advises stopping the task if ChatGPT opens the wrong site or uses incorrect information.
• Actions that commit you. Orders and payments can be hard to undo: in its cloud browser, ChatGPT is designed to ask for confirmation before such actions (OpenAI help page, 20 October 2026).
• Hidden instructions. A web page, an email or a document can contain text that pushes the agent to act against you (prompt injection). Anthropic calls it the biggest risk of AI tools with a browser and says the risk is not zero; NIST lists it among the risks of agent systems.

Agents, data and permissions

What an agent sees becomes part of the task: Anthropic writes that for Claude in Chrome what is visible in the tabs it works in becomes part of the conversation. Under the MCP specification (version 2026-07-28), the app must obtain your consent before exposing data to a server or using a tool, but the protocol cannot enforce this on its own: it depends on the app.

Anthropic advises a separate browser profile without access to banking, health or government accounts. For ChatGPT’s cloud browser, OpenAI advises against allowing access to every website and asks you never to paste passwords, security codes or payment details into the chat. To connect or disconnect a service see the guides to ChatGPT connectors and Claude connectors; for what you type into chats, ChatGPT and personal data; before installing an app with AI, the ten questions.

Groceries and agents: what exists today

The guide to AI shopping agents sums up what they can do in Switzerland today, with the vendor pages read in October 2026: an agent can prepare the list and, on a paid plan, fill a cart, but the final say on payment stays with you.

To start, you do not need an agent: keep a list of your own and decide yourself when to pay. Rappn, a free Swiss app (iOS and Android) that compares the current promotions of seven Swiss retailers and turns receipts into a spending overview, helps with exactly this task: the list shows an estimated price (prices are indicative), and the promotions are updated daily as the chains publish them. Rappn covers groceries only, not rent, insurance or taxes. Rappn is building an AI assistant for groceries: you can join the waitlist. A Rappn connector for AI assistants is in development.