Skip to main content
The Playground lets you run the same question against multiple configurations at once, so you can see which model, temperature, and instructions combination answers best before you make it live. Open Agent → Playground from the agent sidebar.
The Playground is a desktop feature. On smaller screens you’ll be asked to switch to a larger display.

​
Availability

  • Advanced Agent only. The Playground requires the v2 Advanced Agent. If an agent is on Simple RAG, the Playground prompts you to switch it first. Newly created agents open on the Sandbox rather than the Playground.
  • Owners and Admins. Members with the User role don’t have Playground access.
  • Desktop only (see the note above).

​
Prerequisites

​
Compare configurations

The Playground gives you side-by-side configurations — Configuration A and Configuration B — each with its own model, temperature, and instruction version, running in parallel.
1

Open the Playground

In the agent sidebar, open Agent → Playground.
2

Set up each configuration

For Configuration A and Configuration B, pick a model, adjust temperature, and choose the instruction version to use. Need a third? Select Add Configuration.
3

Ask one question

Send the same question to every configuration and compare the answers side-by-side.
4

Apply the winner

Once you know which configuration wins, select Apply to Live Agent to make it your agent’s live setup.
The Playground comparing Configuration A and Configuration B side-by-side.
Changes here don’t affect your live agent until applied — experiment freely, then use Apply to Live Agent when you’re ready.
Playground message sends are free. Unlike the Sandbox, which runs a real chat session and uses credits, comparing configurations in the Playground doesn’t consume credits — so it’s the place to try as many models, temperatures, and instruction versions as you like at no cost. See Credits.

​
Voice input

Each configuration’s input box lets you type your question or use Start recording to speak it — handy for testing how the agent handles natural, spoken phrasing. Recordings can be up to 30 seconds and are transcribed into the input box before they’re sent.
In the live widget, each successful voice message costs 1 credit for the transcription — see Let visitors speak their questions. Playground comparisons remain credit-free as described above.

​
Tips

  • Change one variable between configurations (e.g. same instructions, different model) so you can attribute the difference to that variable.
  • Test your trickiest real-world questions here — the Playground is where configuration decisions are cheapest to make.
  • Playground testing is for iteration and doesn’t represent your live visitor experience; final-check in the Sandbox preview.

Sandbox

Model & temperature

Instructions