Skip to content
Docs/AI Control Centre/Test an agent safely and compare prompts

Test an agent safely and compare prompts

A sandbox that records nothing, and an A/B prompt test with a side-by-side comparison.

Staff portalWho: Administrators and academic managersLast updated 3 September 2026

Sandbox

  1. Open AI → More → Sandbox. Select an active agent and, where relevant, a Programme Context so the test includes the programme's safe metadata.
  2. Type a harmless, explicit test question and Send. Open the trace: it identifies the sandbox execution, response time, token use, agent, model and programme context.
  3. Clear Chat empties the browser-only transcript.

Prompt testing

  1. Open AI → More → Prompt Testing. Select the agent; its current instructions pre-fill Prompt A.
  2. Enter the test message and an alternate system instruction in Prompt B, then Run Test.
  3. Read the responses side by side and the Comparison Summary: tokens, time difference, the faster prompt.
NoteBoth tools are deliberately non-persistent. No session, message, memory, interaction, audit or action record is created.

Related guides

PreviousSet AI approval policies
Was this guide helpful?