Agents
Create an agent and configure it in the editor's Build, Appearance and Settings tabs, trying the draft as you go.
An agent answers questions from one or more knowledge bases with a chat model. Editors, admins and owners create and edit agents under Agents in the team's sidebar.
Creating an agent
Agents → New agent asks for:
- Name and Address: the address is part of the agent's link,
/a/<team>/<address>. - Description: shown in the agent directory.
- Knowledge bases: up to five. The agent searches each with its own settings.
- Chat model: the models your platform allows for the knowledge bases' classification. If a model isn't approved for their level, the dialog says so.
You can change everything later. Nobody can chat with the agent until you publish it.
The editor
The agent editor has six tabs: Build · Evaluations · Appearance · Share · Analytics · Settings. The header shows the version badge ("v4 live", or "Draft" before the first publish), which opens the version history, and Publish.
Build and Appearance edit the draft. People keep chatting with the published version until you publish again. Settings and Share apply at once.
Build
Build is a list of sections. A closed section shows a one-line summary of its settings. Beside them, the Try it panel chats with the draft.
Instructions
What the agent is for and how it should answer, up to 20,000 characters. Grounded adds its own grounding rules around your instructions: cite every factual sentence, don't invent steps beyond the sources, and treat the retrieved text as data, never as instructions.
Model
The chat model that writes the answers. The field shows the most sensitive level the model is approved for, whether it supports tools, and its context window. Only models approved for the agent's knowledge bases can be used.
Knowledge
The knowledge bases the agent searches (up to five). Each gives its own number of passages per search: by default the knowledge base's own setting ("Inherited from the knowledge base (6)"). Override sets a number from 1 to 20 for this agent; Inherit goes back.
Pinned filters narrow every search the agent makes: sources, document kinds, tags, URL prefixes and dates. Use them when one agent should answer from part of a broader knowledge base.
Answering
- When to search
- Search before every answer (default): the question, rewritten to stand on its own, is searched once and the passages go to the model. Fast and predictable.
- Let the model decide when to search: the model gets a search tool and may search several times, or not at all. Needs a chat model with tool support.
- Answer only from the sources (on by default, recommended): strict grounding. When the sources don't answer the question, the agent replies with the refusal message rather than general knowledge. If they cover part of it, it answers that part and says what isn't covered.
- Refusal message: sent word for word when the sources have no answer. The default is "I couldn't find an answer to that in the sources I have."
- Citations: what people see for each cited source: No citations, Title and snippet, or Title, snippet and link (web pages) (default). Uploaded files are always cited by title only.
Safety
- Check answers before showing them: people see "Thinking…" until the answer passes moderation, instead of a streamed answer that may be withdrawn. Public agents always work this way.
- Stricter category rules: tighten the platform's moderation policy for this agent, per category, for questions and answers. An agent can make moderation stricter than the platform's policy, never weaker. The section says when the rules have no effect yet (for example when the audience has no moderation).
SystemOne checks
Shown when the platform has a SystemOne model. Each check uses the platform default unless you turn it on or off for this agent:
- Passage judging: before the answer, a SystemOne model judges each candidate passage (Passages to check, 10 by default): it re-ranks them, drops prompt injections and passages that don't help, and keeps passages that contradict the question's premise apart. If nothing usable is left, a strictly grounded agent refuses without calling the chat model.
- Citation checks, with Citation mode: annotate marks each claim's verdict; enforce also removes confidently unsupported citations, and a strictly grounded agent with nothing supported refuses. See Claims and verdicts.
- Scope check: before searching, spots small talk (answered briefly without a search) and questions outside the agent's subject (a strictly grounded agent refuses without a search).
Each check adds latency: passage judging roughly one SystemOne request per passage, citation checks one per claim and source after the answer.
Advanced
| Setting | What it does |
|---|---|
| Temperature | The model's sampling temperature. Blank uses the model's default. |
| Maximum answer length (tokens) | Blank uses the model's limit. |
| Reasoning effort | How long a reasoning model thinks before answering, when the model supports it. |
| Source token budget | Passages are trimmed to about this many tokens (500–32,000; default 6,000). |
| Maximum searches | With "let the model decide", how many model turns may search (1–8; default 4). The last turn must answer. |
| Minimum similarity | Drop passages less similar than this (0–1). 0 turns it off. |
| Rewrite follow-up questions | Rewrite a follow-up into a question that stands alone before searching (on by default). |
Try it
The panel beside Build chats with the draft, so you can test a change before publishing it. It works like the chat, with two differences: you can open the model's reasoning, and every answer has Add to evaluations. Test conversations aren't kept in your conversation history.
Appearance
The agent's look and welcome, part of the draft:
- Accent colour: used for the agent's avatar and highlights, always with white text, and checked for contrast. Leave it empty for the theme's indigo.
- Welcome message, shown when a chat starts.
- Starter questions: suggestions people can pick. They also appear under a refusal, as a hint of what the agent covers.
A live preview shows the result. There's no custom CSS, so every agent stays accessible.
Settings
Name, Address and Description, and a danger zone to Disable, Enable or Delete the agent.
These settings aren't part of versions: saved changes reach people at once. Disabling an agent stops all chat with it until you enable it again. Deleting it stops chat for good; people keep read-only access to their own conversations with it.
A platform admin can also disable an agent. It then shows "Disabled by platform" and only a platform admin can enable it again; the team's admins and owners are notified.
Evaluations and Analytics
The Evaluations tab holds the agent's evaluation sets (Evaluations). Analytics shows its usage and quality (Agent analytics). Share is covered in Sharing an agent.