Test in Playground

Open Playground from an agent’s sidebar to test its saved knowledge in an authenticated workspace session. The normal view starts from the agent’s saved model and system prompt. It is the safest place to compare a candidate configuration before exposing it through a public widget or channel.

Test without saving

Choose a Model or edit the System Prompt, then select Apply & Test. This remounts the test chat with the current overrides but does not persist either value. The Unsaved changes badge means the editor differs from the saved agent. Select Save to Agent only after the candidate passes your acceptance set; Reset restores the last saved model and prompt.

Keep a prompt within 8,192 characters if you expect to save it. The chat endpoint accepts a longer test prompt, but the saved System Instruction has the lower limit. A normal reply can show expandable source chips after streaming when retrieval produced sources. They are useful evidence, but they do not appear for every answer and should not be treated as guaranteed citations.

Compare models

Select Compare Models to begin with two panels: the saved primary model and Claude Sonnet 4.6. You can use two through four panels. A shared input sends the same message to every unpaused panel and shows the summed per-message credit cost first.

Each panel can vary model, temperature, and prompt. Its configuration is stored locally in the browser for that agent, not as shared server state. After a reply, the panel shows timing, token, and credit statistics. Save a panel only after confirming: it overwrites the agent model and system prompt, while temperature remains Playground-only.

All comparison panels share the agent’s saved sources and retrieval settings. Use normal mode when you need source chips to diagnose retrieval, and use the response-quality guide to grade results consistently.