What is the difference between AI agents and chatbots

A chatbot can answer a question or help draft something in conversation. An AI agent can also work through steps using tools, such as finding information and preparing an action in another service.

The names overlap. A product can look like a chat window while also operating a browser. Check what it can do rather than trying to identify an agent by its appearance.

What happens after the answer?

Suppose you want somewhere for a team lunch. A simple chatbot interaction might produce suggestions from the information you provide. You check the venues and make the reservation yourself.

An agent could search current listings and use what it finds to build a shortlist that fits your requirements. If you ask it to continue, it might then prepare a booking. It can choose a next step based on the results, rather than waiting for you to describe every step.

Anthropic’s explanation of agent patterns distinguishes predefined workflows from systems that decide more of their path. It also acknowledges that agent has several definitions. Treat the label as an introduction, not a complete description of the product.

Does an agent always take action?

No. A system can use tools to research a question and return a report without sending anything or purchasing anything. Tool use and authority to make a commitment are different capabilities.

Ask for the outcome you need. If you want options, say so. If you want a draft enquiry, make clear that it is for review. Before a booking, confirm the exact offer you are accepting.

A useful request might say the following.

Find three venues for ten people on my chosen date. Compare the current offers against my budget. Do not contact them or book. Show what still needs confirmation.

That is a real task with a clear boundary. You can expand it later without granting broad authority at the beginning.

What extra access does the task need?

An agent does not inherently need your whole inbox. Public research can use public sources. Preparing a personalised meeting brief might need a relevant calendar event or an authorised document.

Read connection requests against the task. Permission to read information is different from permission to change it. Read-only access still has privacy implications because the assistant receives information.

If the source contains someone else’s confidential material, your ability to open it does not automatically authorise sending it to another service. Use the approved environment for work information.

Why does review become more important?

A wrong answer can mislead you. A wrong external action can also send a message to the wrong person or commit money before you notice.

External material can contain instructions intended to redirect the assistant. NCSC explains prompt injection as a system-design challenge, so protection cannot rest entirely on you spotting every suspicious sentence.

Use the product’s supported permission controls. Before a consequential action, inspect what will actually happen. A vague continue prompt gives you less to judge than the exact recipient and message.

Afterwards, check the relevant confirmation. A claim that a booking is complete should correspond to a real booking record, not just a fluent sentence.

Where AI Passport fits

Both conversational assistants and agents can benefit from relevant context. AI Passport holds selected personal context that connected assistants can use with your permission.

For a team lunch, you might save a general preference for venues near public transport. The attendee count for this particular meal can stay in the request.

Ask a supported connected assistant to propose the durable preference, then review its wording. A permitted assistant can later retrieve it for another outing. Approval of the memory is separate from giving an app access.

The preference helps define a suitable result. It does not authorise booking or payment. Those actions need their own appropriate authority in the tool carrying them out.

Check whether the particular agent supports a documented Passport integration. If it does not, a short manually supplied brief can still carry the relevant preference, but it creates a copy in that service.

Which should you use?

Use conversation for a question or draft when that solves the problem. Consider agent features when collecting information or coordinating steps is the work you want help with.

Try a bounded task first. Judge the useful result after correction and checking. More actions are not automatically more value, particularly if you spend the afternoon repairing them.