The Playground lets you talk to an agent using its latest saved configuration, whether or not that revision is published. Use it to try changes before making them available in your installed assistant.

## Try the draft

Open the agent's **Playground**. Its configuration and chat sit alongside each other, so you can edit the [prompt](/prompt), [model](/model), or other agent settings and then send a message.

The Playground saves pending configuration changes before sending. Each new turn uses the latest saved revision. Saving while a reply is running does not change that reply; the updated configuration applies on the next turn.

Testing uses the agent's configured knowledge scope and published skills. It does not make unreleased documentation or unpublished skill instructions available. See [Agent knowledge](/agent-knowledge) and [Skills](/skills).

## What needs your installed app

The Playground can show how the agent responds, uses its accessible documentation and skills, asks follow-up questions, and requests approvals. It is not an end-to-end test of your frontend integration.

Page reading, navigation, highlighting, form prefilling, and registered API action execution need the assistant installed in your app. The Playground has no connection to the customer's page or your actions adapter, so it cannot execute those requests. Test them through the installed assistant with your app's navigation callback and [adapter](/adapters) configured.

See [Install](/install) for the integration, [App access](/app-access) for page tools, and [Actions](/actions) for API operations.

## Publish the configuration

When the draft is ready, [publish a revision](/revisions) to make its configuration available to users. Playground messages do not publish it.

For missing information or unexpected answers, see [Agent troubleshooting](/troubleshooting-agents).