Creating an agent

An agent is a persona, a model, and a focused set of knowledge. It reads only the sources you scope to it, and each person who asks still sees only what they are already allowed to see.

Before you start

You need the admin or owner role. Agents only shows in the menu for admins and owners, and other roles are sent back to chat.
Open the user menu at the bottom of the sidebar, then choose Agents under Manage. If your organization has no agents yet, a Default Agent is created automatically the first time an admin opens Agents or chat.

Start a new agent

Click New agent at the top of the Agents page, or the Create an agent tile at the end of the grid.
On the first step, pick a starting point under Start from a template: Operations analyst, Customer support, Policy reviewer, or Blank. Templates are editable defaults, nothing more.
Move through the seven steps with Back and Next (the button reads Review on the step before the last). The footer shows "Step N of 7" and the step's name.

Step 1: Identity

Name is required. The agent's permanent handle is derived from the name when you create it and never changes; there is no handle field.
Public name is what visitors see on the widget. Description is one line shown on the agent card.
Choose a look: Icon, Emoji, or Image (PNG, JPG, WebP or GIF, 2 MB max).
The Active switch controls answering. Inactive agents stay editable but stop answering on every channel.
Persona is who the agent is and how it sounds. Pick a tone under Start from a tone, write Voice & personality, or click Draft a persona from the name & description. Starter prompts for internal chat can be Auto-generate or Custom.
Instructions is standing guidance carried into every answer on every channel. The agent follows it, but nothing checks the answer against it afterward. For a limit that is enforced, use the Boundaries step.

Step 2: Behavior

Model: choose Single model and pick one from the searchable list, or choose Council to run several models on every answer and merge them. See Council answers.
Answer style: set Depth (sources) (Auto, Off, Fast, Standard, Thorough, Research), Effort, Creativity, Answer length, Cited figures, and whether the agent may look at cited images (Figure inspection).
How it answers: Strict answers only what it can cite, Balanced prefers cited evidence but may bridge small gaps, Flexible treats knowledge as context rather than a fence.

Step 3: Boundaries

Who can use it: Everyone, Specific people (pick teams, roles or people, each set to Allow or Exclude), or Just me. With Specific people, admins and owners can chat with it without a grant. With Just me, other admins can still see and manage it, but cannot chat with it.
What it may read & reveal: the Sensitivity ceiling (Standard, Confidential, Highly sensitive), PII in answers (Redact or Allow internally), and Most it may link to.
Answer rules: turn on Enforce answer rules, then choose rules such as Citable answers only, Never recommend competitors, or Decline & redirect off-topic, and add Blocked topics.

Step 4: Knowledge

Choose Everything, Selected knowledge, or No knowledge. With Selected knowledge you add sources, items, tags, categories and teams as include or exclude rules. Public web search is on by default and can be limited to certain sites. See Knowledge scope and audiences.

Step 5: Skills

Turn on what the agent may do beyond answering. Skills are grouped as Platform, Custom and Connectors, and each carries a risk label: Read-only, Drafts, Takes action, or Admin. On a saved agent these toggles save instantly. See Skills.

Step 6: Channels

Internal chat is always on. Toggle Public widget, Email, API, MCP server, and Slack as needed. The Slack tile stays off until Slack is connected.

With API or MCP server on, the step shows the handle that callers use, with a Copy button.
With Public widget on, the widget's look, greeting, lead capture and handoff routing are set in Help Desk → Widgets once the agent is saved. The step shows a Widget editor link as soon as the agent exists. See Website chat widget.
Email inboxes pick their answering agent in Help Desk → Channels. See Email inboxes and forwarding.

Step 7: Review and set live

The Review step lists each section with an Edit link, or Finish when something still needs attention, and tells you when the agent is ready: "Ready to publish — every section is set."

Set live (or Save & set live on a live agent) publishes the agent across its channels. If the widget is on, Set live & set up widget takes you straight to the widget editor.
Save draft keeps your edits and sets the agent inactive. On a live agent this takes it offline.
If the knowledge scope is empty you'll see "Set live with nothing in knowledge scope?" You can add an include rule or choose Set live anyway.

Test the agent

Click Test agent in the builder header, or Test on the Review step. The side panel is a private sandbox: it runs real retrieval and policy with your current edits, writes no analytics, and never affects production. Each answer shows a confidence badge, the number of citations, and "Would hand off" when a handoff would have been suggested. You can change the effort per question.