Chatbot, assistant or agent: which one is it
A chatbot answers what you type. An assistant carries a task across several turns and keeps the goal in view. An agent takes actions in the world without asking each time. Most products called agents are assistants, and most called assistants are chatbots.
The 3 words are used as if they were fashions rather than categories, and the difference between them is the amount of harm a mistake can do. Reading a product page for which of the 3 it really is takes a minute and saves an hour.
A chatbot answers a turn
The unit of work is one message. You write, it replies, and the exchange is complete. Memory of earlier messages helps the reply read naturally, and it does not change what the unit is. Nothing happens outside the conversation, which is why a mistake costs you a re-read rather than anything else.
An assistant carries a task
The unit of work is a goal that spans several turns: draft this, then shorten it, then adapt it for another audience. The system holds the goal while the wording changes. This is where memory stops being a nicety, because losing the goal in the middle wastes everything before it.
An agent acts
The unit of work is a change in the world: a message sent, a file written, a purchase made. Autonomy is the whole point and it is also the whole risk, because a wrong step is not undone by asking again. An agent that asks permission for every step is an assistant with extra latency.
Which one you are using
Look at what happens when the system is wrong. If you re-read and rephrase, you have a chatbot. If you lose a piece of work, you have an assistant. If something outside the screen changed, you have an agent, and the safeguards should be proportional to that.
Before you paste it in
Is a chatbot less capable than an assistant?
Not at the level of a single answer. The same model can sit behind both, and the difference is how much the product lets the model carry between turns.
Why do so many products call themselves agents?
The word signals capability and it costs nothing to use. Checking what the product actually does outside the conversation settles it faster than reading the page.
Does an uncensored chatbot become an agent if it can browse?
Browsing adds a source of information rather than the ability to act. Reading a page is not the same as changing something, and the risk profile stays closer to a chatbot.
Open a chatbot that answers turns and does nothing outside the tab.
Open uncensored chatbot