> ## Documentation Index
> Fetch the complete documentation index at: https://docs.weav.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Playground

> Test your AI agent’s real response behavior safely.

## What the Agent Playground is

The **Agent Playground** is a built-in testing panel on the agent settings page that lets you ask your agent questions and see how it responds in real time.

It's main purpose is to **validate agent behavior before you deploy it to your customers**.

***

## Why teams use the Playground

Playground helps you:

* Test answers without embedding the chat on your site or sending emails
* Validate training data quality before launch
* Check if tone, response length, and prohibited phrases are properly configured
* Confirm escalation behavior in a safe environment
* Catch weak or hallucinated answers early

<Info>
  Think of Playground as your “staging environment” for agent quality.
</Info>

***

## What is different from production

Playground is **safe and non-production**. It does **not** behave exactly like a live customer conversation in every way.

### Not saved to inbox

Conversations in Playground are **not** written to your inbox as customer conversations.

### No customer-facing side effects

Playground tests do not send real customer emails or create live chat messages.

### Session is temporary

Your test thread exists in the sidebar session and can be reset using the refresh control or refreshing your browser.

***

## Best ways to use the Playground

### 1) Smoke test before publishing changes

After updating instructions, context, training, or prohibited phrases:

* Ask 5–10 representative customer questions
* Confirm answers are accurate and on-brand
* Verify sources look relevant

### 2) Test edge cases

Try hard prompts such as:

* Ambiguous questions
* Policy exceptions
* Escalation-sensitive requests
* Out-of-scope questions

### 3) Run “before vs after” checks

Use the same test prompts before and after major changes to see whether quality improved.

### 4) Verify source quality

If response sources look weak or unrelated, improve your training data before deployment.

***

## Tips for non-technical users

* Start with your top 10 real customer questions
* Keep a reusable prompt set so testing is consistent
* If output quality drops, review training data first
* Save settings before testing if you see an “unsaved changes” notice (so preview reflects latest config)