Create your first agent

This tutorial takes you from the Agents dashboard to a tested agent. Choose a source whose contents you can verify yourself; that makes it possible to tell whether the agent is actually grounded instead of merely sounding convincing.

1. Create the agent

Sign in at agentstack.build and open your workspace. The dashboard lists its agents.

You need permission to create agents in the selected workspace. The full walkthrough also requires permission to add sources and edit the agent; viewing the final conversation in Activity requires Activity access. Confirm the organization shown in the workspace switcher before continuing, because agents and their knowledge belong to that organization.

1. Select New AI agent. In an empty workspace, you can also select Create Agent in the empty state.
2. In Create New Agent, enter a Name such as Customer Support Agent.
3. Select Create Agent.

AgentStack creates the agent and opens its Analytics page. You can return to the Agents dashboard to select the agent later; selecting its card also opens Analytics. If the dialog says that the workspace has reached its agent limit, reduce the active-agent count or upgrade the plan before creating another one. Scheduling an unneeded active agent for deletion removes it from the limit count.

2. Add one reliable source

In the agent sidebar, expand Sources and select Content. The Content page supports Website, Pages, Files, Notion, and Q&A sources.

For a simple first test:

1. Keep the Website tab selected.
2. Under Crawl Website, enter a public URL that contains a fact you know, such as a shipping or returns page.
3. Set Max Pages narrowly enough to include the relevant page without pulling in an entire unrelated site.
4. Select Fetch.
5. Wait for the website status to become Completed. If it becomes Failed, inspect the displayed error before testing the agent.

The website table reports its status and page count. The Pages tab shows the individual crawled and indexed pages. Do not treat a submitted crawl as ready knowledge while it is still Crawling or Processing.

If a website is not the right source, a focused Q&A pair is useful for a single canonical answer. Files and Notion are better when the authoritative material already lives in those formats. The full Sources guide covers the available options.

3. Give the agent a narrow job

Open Playground. The left panel contains a Model selector and a System Prompt editor. Start with a short instruction that covers:

• the audience and job;
• which subjects are in scope;
• the desired answer style;
• what to do when the sources do not support an answer;
• when to ask a clarifying question or offer human help.

For example:

You are the support agent for Acme Store. Answer customer questions using the
available company knowledge. Keep answers concise and do not invent policies.
If the knowledge does not contain the answer, say that you are unsure and
provide the configured support contact or explain the manual next step.

Select Apply & Test to use the edited model and prompt in the Playground chat. This applies the changes to the test session; it does not make them the agent's saved configuration. After the test behaves correctly, select Save to Agent. Reset restores the last saved model and prompt when there are unsaved changes.

A newly created agent has no actions. It cannot transfer a conversation or perform another handoff action until you configure that capability. Until then, the prompt should offer only a support contact or other manual route that exists.

4. Test supported and unsupported questions

Ask at least three kinds of question:

1. A direct question whose answer appears verbatim in the source.
2. A paraphrase that requires retrieving the same fact with different wording.
3. A question that the source does not answer.

Check the facts, not just the tone. A good unsupported response should follow the boundary you wrote in the system prompt instead of fabricating a plausible policy. If a known answer is missing, verify that the source is Completed, that the relevant page appears under Pages, and that the page actually contains the needed text.

Use the response-quality checklist for a more rigorous evaluation.

5. Prepare a first deployment

When the acceptance questions pass, open Settings → General and confirm the saved model, system instruction, and Public setting. A non-public agent does not render its embed widget on external sites.

Then open Deploy. The Chat Widget tab provides the embed script, while Domain Protection can restrict the widget to one hostname per line. Enter bare hostnames such as example.com, without https:// or a *. wildcard. Leave the domain list empty only when you intentionally want to allow all domains. Follow the widget installation guide for placement and verification.

After installing, open the customer site in a private browser window and repeat one supported and one unsupported question. Your first agent is ready for a limited launch when the live widget matches the Playground behavior and the resulting conversation appears in Activity.