Playground
The Playground is where you try a model or an agent before you rely on it. You need the developer or admin role on the project.
Start a conversation
- Open Playground.
- In the target picker, search for and choose an agent under Agents, or a model under Models.
- Type a message and send it.
When you pick an agent, each turn is a real agent run with the agent's prompt, tools and knowledge base. For a code agent you can also pick the version: the live one, the staging one, or any deployed version. When you pick a model, the message goes straight to the model through the Model Gateway and the reply streams in.
Models that are built but not deployed are shown, but you cannot pick them. The picker says why.
To send images, attach, paste or drop them into the message. This works for a vision model, an external provider's model, or an agent whose first step uses a vision model. Earlier turns keep an [image] marker in place of the picture. See Images.
Sessions
Each conversation is saved as a session once you send the first message. Sessions are listed on the left, newest first, and you can search, rename and delete them. Deleting a session also deletes its transcript and tool-call logs.
- Sessions belong to the project. Every developer and admin on the project can open them.
- A session stays with the agent or model it started with. To talk to a different target, start a new session.
- Sessions that have not been used for 30 days are deleted.
Inspect a turn
Choose Inspect on a reply to open the inspector. It shows:
- timing: time to the first token, total duration and throughput
- token counts, as the model reported them
- each tool call the agent made, with its arguments and result
- for a knowledge base search, the chunks that came back, or that no chunk matched
- a link to the full run in Runs & Traces, for an agent turn
A tool call marked as written out as text means the model did not return it in the structured format. It ran, but that model's tool calling is unreliable.
Approve a tool call
When an agent wants to call a tool that needs approval, the turn shows Waiting for your approval with the tool name and its arguments. Choose Approve or Reject, and optionally give a reason. The reason is recorded in the audit log.
If you reject the call, or nobody decides before the approval window closes, the run stops without an answer. See Human approval.
Talk by voice
Choose Switch to voice to talk to the same agent or model through your microphone. Allow microphone access when your browser asks. The spoken turns are saved to the session as you talk.
In a voice conversation, a tool call that needs approval waits on the line until someone decides. See Voice agents.
The inspector shows each voice turn's legs: the speech-to-text delay, the model and the text-to-speech time. See Voice agents for what a voice session needs.
Save a turn as an eval case
Choose Save as eval case on a reply to keep the exchange as a regression test.
- Add checks: text the reply Must say, text it Must not say, or a tool it Must call.
- Choose an eval suite, or create a new one.
- Choose Save case.
The case keeps the conversation before the turn, and records which agent, version and model produced it. Saving the same turn again updates the existing case. See Evals.