Agents

Testing an agent

The preview in the editor is a real chat with your unsaved draft. You test exactly what your team gets once you save.

On the right of the editor, you chat in the Preview with the agent as it stands on screen, unsaved changes included. Until you save, your team's chats keep running on the saved agent.

Test and improve

Start with a conversation starter

Click a conversation starter in the empty preview chat. This shows what your team sees first.

Try real cases

Enter a typical request, then an incomplete one and a case the agent should decline or hand on.

Adjust the draft

Change what does not fit, under Describe or directly under Configure.

Ask again

Choose Edit and resend on your question. The new answer already uses the changed draft.

Save

Click Save changes (for a new agent, Save agent). Only now does the change reach your team.

How the preview works

  • A real chat: the same models, data protection, brains and tools as anywhere else. It therefore costs credits like a chat.
  • Always the current draft: instructions, model and brains are taken over before every message. A file you add or remove arrives in, or leaves, the running test chat at once.
  • Your rights: you test with your access to brains and your own connectors. If security mode applies to you, it applies here too.
  • Not in your history: the test chat appears neither in the chat list nor in search.

Start over and hide

Start over begins an empty test chat with the same draft. After more than 30 minutes without activity the preview also starts empty; reloading the page keeps the conversation. Hide preview gives the fields the full width. On a phone, the preview is a tab of its own.

The “Preview” of the Quote Writer agent with “Start over” at the top, then icon, name, description, the message field and the conversation starters.

What to test

CaseWhat to look for
Typical requestAre content and format right?
Details missingDoes the agent ask instead of guessing?
Edge caseDoes it decline or refer on, as required?
SourcesDoes it name the right documents? Open one source.

Best practices

  • A fixed test set of five to eight real requests, asked again after every larger change.
  • One change at a time, or you will not know what made the difference.
  • Someone from the audience should try the saved agent, since they may have different rights.

Frequently asked questions

Next steps

On this page