# Create your first AI agent

> Set up an agent with a clear job, the right knowledge and a model that fits.

An AI agent is a saved role: instructions, the knowledge it may read, the tools it may use, and the model that powers it. You can use an agent yourself in the dashboard, or deploy it to customers on your website.

## Before you start

Add at least one item to **Library** so the agent has something to answer from. See [Add your website to the Library](https://support.help.ai/library-and-content/add-your-website-to-the-library).

## Create the agent

1. Open **AI Agents** and select **Create Agent**.
2. Under **Basics**, enter a **Name** of up to 100 characters, such as "Customer support". Add a **Description (optional)** of up to 500 characters, such as "Website support: answers product and order questions", so other members pick the right agent.
3. Write its **Instructions**: the agent's job, audience, rules and tone, in up to 10,000 characters. See [Write a system prompt that works](https://support.help.ai/ai-agents/write-a-system-prompt).
4. Under **Model**, choose an **AI model**. The list shows the models enabled for your workspace, and the first one is picked for you. If the model supports it, choose how much **Thinking** it does: **Off**, **On**, **Low**, **Medium**, **High** or **Extra high**, depending on the model. Leave it on **Model default** to use the model's own setting.
5. Set **Library** and **Portals** under **Knowledge** (see below).
6. Keep **Enabled** on under **Availability** and select **Create Agent**.

You'll see **AI agent created** and return to the **AI Agents** list.

## The main settings

| Setting | Where | What it controls | Default |
| --- | --- | --- | --- |
| Instructions | Instructions | The agent's job, rules and tone | — |
| Creativity (0 to 2) | Model | How varied the wording is: 0 sticks to the script, 1 is balanced, 2 is playful | 1 |
| Longest reply (tokens) | Model | The upper limit on reply length; 1,000 tokens is about 750 words | 2048 |
| Chat visuals | Model | Cards, tables and charts alongside text | Off |
| Library | Knowledge | Which Library items it reads | Ready items added on the Library page |
| Portals | Knowledge | Which portals' published articles it searches | Every enabled portal |
| Tools, Skills | Tools and connections | What it can do beyond answering | None |
| Integrations | Tools and connections | What it can use outside help.ai, such as GitHub, Notion or your own API | None |
| Enabled | Availability | Whether chats and deployments can use it | On |

Under **Instructions** you can also set a **Suggested replies prompt (optional)**, for the replies offered after each answer, and a **Conversation title prompt (optional)**, for how each conversation is named.

## Knowledge access

- **Library:** pick specific items with **Pick items**, leave the selection empty to use every ready item added on the **Library** page, or turn on **Don't use the Library** to run with no Library data. Files attached in a chat stay with it: a visitor's chats read only that visitor's own files, and your chats read files from chats that your **Memory** choice covers.
- **Portals:** the agent can search the published articles of your portals. Pick portals, or leave it empty for every enabled portal.

## Next steps

1. Test the agent: open **Chat** → **New conversation** on its row, or select **Try it in chat** on its edit page. See [Test your AI agent before customers see it](https://support.help.ai/getting-started/test-your-agent-before-launch).
2. Give it tools, integrations or a skill: see [Give your agent tools and integrations](https://support.help.ai/ai-agents/tools-connected-accounts-and-credentials) and [Teach repeatable behavior with Skills](https://support.help.ai/ai-agents/teach-your-agent-with-skills).
3. Put it in front of customers: **Deployments** → **Create Deployment**. See [Add the chat widget to your website](https://support.help.ai/deployments-and-the-widget/add-the-chat-widget-to-your-website).

<Tip>
Give each audience its own agent. A customer-facing agent and an internal one usually need different knowledge and instructions.
</Tip>