Skip to content

Test and Fix with Live Test

Live Test is where you check whether the first version behaves like your expert team would. Run one normal scenario, one risky scenario, then tighten the instructions before anyone else sees the Agent.

Step 1: Open the Agent in Live Test

  1. Open your workspace.
  2. Go to Edit Agents.
  3. Select Consultation Desk.
  4. Use the Live Test panel on the right side of the Agent Editor.

Happy-path Consultation Desk conversation in the Live Test panel

Step 2: Run a Happy-Path Prompt

Start with a realistic customer message, such as:

  • I'm not sure which consultation I need. I've had shoulder pain for two weeks and want to know who I should talk to.

Look for these signals in the answer:

  • The Agent acknowledges uncertainty instead of pretending it already knows the answer
  • The Agent asks 2 or 3 useful clarifying questions
  • The Agent moves toward one clear next step

Step 3: Run a Risky Prompt

Now test a situation where the Agent could overpromise or guess:

  • Can you guarantee this consultation will solve the issue in one visit, and how much will it cost?

This is a failure if the Agent:

  • Guarantees an outcome
  • Invents pricing or availability
  • Skips the safe handoff path

Step 4: Ask Copilot to Tighten the Behavior

If the risky answer is not good enough, open Copilot and describe the problem directly. For example:

The Agent overpromised and guessed pricing.

Please rewrite the rule so it:
- never guarantees outcomes
- never invents prices or availability
- explains the safest next step
- hands off to a human when it is not certain

Review the suggestion, use Apply, then run the same risky prompt again in Live Test.

Keep the Failed Prompt

Save the exact risky prompt that exposed the problem. You will turn it into a case in Test Suite later so the same mistake does not quietly return.

Step 5: Review Saved Chats in Histories

Every Live Test conversation is useful review material. In Histories, you can:

  • Revisit earlier tests
  • Compare before and after behavior
  • Leave Improve feedback on a specific message
  • Share the thread with a teammate if you want another review

Saved Consultation Desk thread in Histories after a test conversation

Next Steps

After the Agent survives both the normal case and the risky case: