Testing an agent
Use the test panel to validate an agent's behavior against the current draft before publishing.
Admins only. Members can run and observe; only admins can edit.
The agent builder includes a test panel that runs prompts against the current draft. Use it to iterate before publishing so members never see a half-tuned agent.
Opening the panel
The test panel slides in from the right edge of the editor. It shows a chat composer and a message stream.
What it runs
Test runs use the current draft — not the published version. That includes unsaved edits in the form, identity changes, and tool selections. Each test conversation is ephemeral: it isn't saved to chat history, doesn't update the agent's memory, and doesn't appear in analytics.
Comparing behavior
Open the test panel and send a prompt. Note the response. Edit the instructions or tools, save the draft, then re-send the same prompt. The two responses sit in the same test thread for direct comparison.
When to use the test panel vs. real chat
Use the test panel for tuning prompts, validating tool selection, and catching regressions. Use real chat — after publishing a draft — to validate end-to-end UX: suggested prompts, attachment flow, and thread persistence.