Skip to main content

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
Think of Playground as your “staging environment” for agent quality.

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)