跳到主要内容

Create an Agent

A complete walkthrough for creating, configuring, testing, and sharing your first agent.

What you need​

  • An admin account with access to the Admin Dashboard.
  • The LLM provider you want to use must have its API key configured. If your provider's models are greyed out in the model picker with "API key not set", a super-admin needs to add the key on the Deployment Settings → LLM Provider API Keys page first.

Steps​

1. Go to Your AI Agents​

Open the Admin Dashboard. It lands on the Your AI Agents page by default. You'll see a list of any existing agents.

2. Start a new agent​

Click the + icon button in the top-right corner of the page (there is no text label — look for the blue icon button). A dropdown appears:

  • Create from Template — opens the template dialog. Pick a pre-configured starting point and click Continue to load it. See Templates.
  • Create from Scratch — opens a blank editor.

3. Fill in Basic Info​

In the Basic Info section:

  • Agent Name — required. Short and descriptive; end users see this name.
    • After typing a name and clicking out of the field, the dashboard may generate a system prompt suggestion. Click into the System Prompt field and press Tab to accept it.
  • Text Model — click the model selector to open the dropdown. Use the search box to filter. See Pick the right LLM.
  • Speech Model — leave as "No speech model selected" unless this agent needs voice support.
  • Thinking Budget — this section only appears when you select a thinking-capable model (e.g., a Claude model). Leave it on Off for your first agent.
  • System Prompt — write the agent's instructions. See System Prompt for tips and a template.

4. (Optional) Add a tool​

Skip this step for your first agent — you can add tools later.

To add one now, in the Tool Integration section:

  • Click the + button → Pre-defined Tool to pick from the catalog, or From Scratch to add a new remote URL.
  • Fill in the fields and save the tool entry.
  • Note: after saving the agent, you still need to go to the Tool Configuration page to authorize and fetch the tool list. See Step 7.

5. (Optional) Add a Knowledge Base​

The Knowledge Base Integration section is always visible in the editor with inline form fields. You don't need to click a separate button — just fill in:

  • Knowledge Base Name and Description.
  • Click Add Documents to upload files or add web URLs.

Leave these blank if you don't need a KB yet. See Set up a Knowledge Base for full details.

important

Each agent supports one Knowledge Base only.

6. Save the agent​

Click Create Agent at the bottom.

If validation fails, you'll see an inline error:

  • "Agent name is required" — fill in the Agent Name field.
  • "Please save or cancel the … tool" — you have an unsaved tool entry. Save or cancel it first.
  • "Invalid JSON" — you're in Advanced mode and the JSON has a syntax error.

7. Finish configuring tools (if you added any)​

Saving the agent is step one of two for remote (MCP) tools:

  1. Go back to Your AI Agents.
  2. Find the agent's card → click View more → Configure Tools (wrench icon).
  3. On the Tool Configuration page: click Configure to authorize (if OAuth-protected), then fetch the tool list and enable the tools you want.

See Add an MCP tool for the full walkthrough.

8. Watch the Knowledge Base status (if you added one)​

The KB status chip in the editor moves through: Setting Up → Configured (or Failed). The agent can only answer from the KB once it reaches Configured. This takes a few minutes depending on how much content you added.

You can also monitor progress on the Tool Configuration page alongside remote tools.

9. Test the agent yourself​

Before sharing with anyone else, open the Share dialog:

  1. On the agent's card → click Share.
  2. Click Start chatting — the User Dashboard opens in a new tab.
  3. Send a few messages. Confirm:
    • The agent's tone matches the system prompt.
    • Tools are called as expected.
    • KB answers reference the right content.
备注

For Start chatting to work, your admin account must also have end-user access on this deployment.

10. Share with others​

From the same Share your ESP Private Agent dialog:

  • Click Copy link → paste into Slack, email, or a doc.
  • Use the QR code → for phone or device handoff.

See Share an agent for access requirements and troubleshooting.

Common gotchas​

SymptomCause
Save fails with "Agent name is required"Agent Name in Basic Info is empty.
Save fails with "Please save or cancel the … tool"An unsaved tool entry exists in Tool Integration.
Can't switch from Advanced to BasicThe JSON has a syntax error. Fix it first.
Model shown in redThe model doesn't support a capability the agent needs (tools, vision). Pick a different model.
Thinking Budget doesn't appearYou're using a model that doesn't support thinking (e.g., Amazon Nova). Switch to a Claude model to see this option.

What's next​

On this page