Agents

An agent is a single chatbot. You can run several — one per site, course, product, or audience — each with its own prompt, knowledge, look, and limits.

The system prompt

The system prompt is the agent's standing brief. It is read before every conversation and shapes tone and scope. Good prompts are specific:

State who the agent is and who it serves ("You help students of Intro to Optimisation").
List what it should do, and what it should refuse or defer.
Set the tone (concise, formal, friendly).

You do not need to tell the agent what language to use. It detects the visitor's language from the question and answers in kind.

Choosing the answer quality

On most plans you pick a quality tier rather than a raw model:

Base — fast and economical, the right default for most assistants.
Premium — a stronger model for harder reasoning. A Premium reply draws more from your monthly message-credit allowance than a Base reply.

Each tier shows the name of the model it currently uses, and a tier with extended thinking is marked as such. On the BYOK and Custom plans you instead pick the exact model and sampling settings yourself.

Reasoning

Some models can "think" before answering. On BYOK and Custom you enable it per agent with Enable reasoning when the model can think. Some models always think (for example Magistral and GLM); for them the switch decides whether that thinking counts as on and can be shown as Thinking. On Base and Premium plans the Quality tier decides.

What visitors see while the assistant works is set by Show the assistant's work on the Configuration tab:

Subtle (default for new agents): one quiet line with the current step, such as searching the documents, running code or searching the web. Nothing stays after the answer.
Tool steps: every step with its duration and search terms. Once the answer starts streaming, the steps fold into a summary such as "3 steps · 4 s".
Thinking: the model's reasoning plus every step, including code it ran, its output and web hits. You can pick it only when the agent thinks: on Base and Premium plans, when the selected Quality tier has extended thinking; on BYOK and Custom, when the model can think and Enable reasoning is on. If an agent that had Thinking moves to a tier or model that does not think, it is saved as Tool steps.

In every mode a "Still working" line appears when no new text has arrived for 3 seconds. All of this is shown outside the answer bubble, and sources are listed under the answer, one per line, following your Show citations setting.

The steps of each reply are kept with it, and you can review them in the conversation detail under "steps". For a web search they include the address and title of up to ten results. Only you see those; your visitors don't. In private mode only the type and duration of each step are stored.

Citations and moderation

Citations — when an agent is grounded in a library, it can show the sources behind each answer. Toggle this per agent.
Moderation — a built-in safety shield is always on. You can opt into additional categories for stricter filtering.
Analyze conversations — optionally have the platform review finished chats for helpfulness and knowledge gaps; the aggregate signals appear in Analytics and the conversations that fell short collect in Knowledge gaps. See Conversation analysis.

Managing agents

Edit an agent any time; changes apply to new messages immediately. Deleting an agent removes it and stops its embed from working, but never touches your conversation history or libraries.