The playground is a private conversation with an agent. Nothing you send reaches a customer, and nothing the agent says is delivered anywhere. It is the place to find out how it behaves before finding out in production.
What to test #
The obvious question
The one the agent exists to answer. If this is wrong, nothing else matters.
The out-of-scope question
Ask about something it should refuse. A good agent declines and offers a human; a bad one improvises.
The hostile question
Try to talk it into a discount, a promise or a policy it cannot make. Better to find that here.
Reading a run #
-
Send a message
Type as a customer would — including the typos and the half-sentences. -
Read the reply and the tool calls
If the agent has tools, you can see which it invoked and what came back. An agent answering confidently without calling the tool that holds the answer is the thing to catch. -
Adjust and re-run
Jump to Configure, change the prompt, come back. Short loops beat long ones. -
Check the audit log
Every playground run is recorded like any other inference, with tokens and latency. See Audit logs.