# Test your AI agent before customers see it

> A short checklist for trying your agent in the dashboard before you install it anywhere.

Customers judge your agent on its first few answers. A short round of testing catches the obvious gaps before anyone else sees them. Test in two places: a test chat in the dashboard for knowledge and instructions, and the deployment's **Preview** for the full visitor experience.

## Start a test conversation

1. Open **AI Agents**.
2. On your agent's row, open **Chat** and choose **New conversation**. Or select **Try it in chat** at the top of the agent's edit page.
3. Find earlier test conversations under **Chat** → **Conversations**.

The agent must be **Enabled** to chat. These conversations stay inside your workspace. They don't appear in **Inbox**, which lists visitor conversations from your deployments.

## What a test chat uses

A test chat in the dashboard runs the agent exactly as you set it up: its **Instructions**, **Library**, **Portals**, model, tools, skills, **Creativity** and **Longest reply**. It has no **Chat settings** of its own. To see the agent in the widget, with the deployment's own settings, use the deployment's **Preview**.

## What to test

1. **Your most common questions.** Use the words customers use, typos included.
2. **Edge cases.** A refund outside the policy window, an order that hasn't arrived, a product you don't sell.
3. **Questions it shouldn't answer.** It should say it doesn't know rather than guess. If it guesses, add a rule to the agent's **Instructions**.
4. **Requests that need a person.** Check that it offers a person instead of making promises.
5. **Tools.** Ask for something that needs each tool you rely on, and check its approval requests. To see them the way visitors do, use the deployment's **Preview** (below).

## Fix what you find

| What you see | What to change |
| --- | --- |
| A wrong or vague answer | Add a **Q&A** or **Text Snippet** to Library with the exact answer |
| An answer from the wrong document | Narrow the agent's **Library** (under **Knowledge**) to the right items |
| The wrong tone or format | Adjust the agent's **Instructions** |
| A procedure it gets wrong every time | Write a **Skill** with the steps |

After you select **Save changes**, the agent uses your changes from its next reply, so you can keep testing in the same conversation.

## Try the full chat experience before you install it

Website widget and hosted page deployments have their own page at your-link.help.chat. Select **Open** on the deployment's row in **Deployments**, or **Preview** (**Open page** for a hosted page) at the top of the deployment. There the agent runs with its own model, tools and skills, the way visitors get it. It's an easy way to test on your phone or share with a colleague before the widget goes on your website.

<Tip>
Leave **Live** on while you test with **Preview**. Nothing appears on your website until you paste the install snippet. Turn **Live** off under **Availability** to stop a deployment answering without deleting anything.
</Tip>

## Related

- [Write a system prompt that works](https://support.help.ai/ai-agents/write-a-system-prompt)
- [Teach repeatable behavior with Skills](https://support.help.ai/ai-agents/teach-your-agent-with-skills)
- [Share a hosted chat page](https://support.help.ai/deployments-and-the-widget/share-a-hosted-chat-page)