Skip to main content
Generic chat answers questions with whatever the model has been trained on. A purpose-built agent answers with your organisation’s knowledge, in your tone, scoped to one job — “Product support”, “HR policies”, “Sales enablement”. This tutorial takes you from an empty agent page to a live, versioned agent your team can pick in chat. You need Editor access or higher. Feature reference lives at Agent concepts and Create an agent; this tutorial stitches those steps into a concrete outcome.

​
Step 1 — Decide what the agent is for

Before clicking anything, write one sentence: “This agent answers X using Y, and does not do Z.” Example: “This agent answers product-support questions using the Help Center folder, and does not give legal or billing advice.” That sentence becomes the backbone of your system instructions — without it, the agent drifts.

​
Step 2 — Create the agent

Navigate to Agents in the sidebar and click New Agent. Give it a Display Name (“Product Support”) and an Internal Name — a URL-safe slug used in API calls and the chat URL (product-support). Add a short description, then click Create. You land on the configuration page. Leave all tabs at their defaults for now.

​
Step 3 — Write the instructions

Open the Instructions tab. Paste a system prompt built from the sentence in Step 1. A reusable skeleton: 
You are the <role> for <organisation>.

Your job is to <task>, using <scope of knowledge>.

Rules:
- Always respond in the user's language.
- Cite the source document when you answer from the knowledge base.
- If a question is out of scope, say so and suggest where to ask.

Tone: <tone>.
Format: <format>.
Pick a Model preset (Fast / Standard / Advanced) that matches the task — Fast is fine for short lookups, Advanced for multi-step reasoning. See Agent concepts — Model for the mapping. Changes save automatically; an indicator in the top-right shows the state.

​
Step 4 — Scope the knowledge

Open the Knowledge tab. Uncheck everything the agent should not read and keep only the folders that match its job. A narrow scope is almost always better than a broad one — fewer irrelevant search hits, shorter context, sharper answers. See Agent concepts — Knowledge. If the folders do not exist yet, create them in the knowledge base first, then come back.

​
Step 5 — Turn off tools you do not need

Open the Tools tab and disable anything the agent should not use. A support agent probably does not need web search. A research agent probably does not need the billing integration. Fewer tools means fewer surprises in production.

​
Step 6 — Add a conversation starter

Open the Conversation starters tab and add two or three example prompts. They appear on the empty-state screen when a user opens a new conversation with the agent, and they also act as built-in smoke tests for Step 7.

​
Step 7 — Test from chat

Open Chat, pick the new agent in the agent selector, and try each conversation starter plus one or two ad-hoc questions. Watch for:
  • Does the agent cite the right documents?
  • Does it refuse out-of-scope questions cleanly?
  • Does the tone match what you wrote in the instructions?
Iterate on the Instructions tab, then retest. This loop is the bulk of agent building.

​
Step 8 — Publish a version

Every edit creates a draft; the live version keeps serving chat until you publish. Once you are happy, click Publish in the version header. Future edits start a new draft — users keep hitting the published version until you publish again. See Agent versions for rollback.

​
Next

Last modified on April 20, 2026