Test and improve your agent

Open the test panel

1. Open Agents.
2. Open the agent you want to review.
3. Select Studio.

The workspace shows configuration on one side and the test chat on the other. This session is temporary and does not create saved conversations.

Prepare representative questions

Do not stop after one simple question. Include:

• Frequent questions the agent must resolve.
• Differently worded versions of the same question.
• Cases with missing information.
• Requests the agent must reject or route elsewhere.
• Questions not answered by the available knowledge.
• Multi-message conversations to check context.

Evaluate each response

Check that it:

1. Resolves the user's intent.
2. Uses available information and does not invent facts.
3. Maintains the expected tone and detail.
4. Follows instruction restrictions.
5. Uses tools only when necessary.
6. Explains the next step when it cannot complete the task.

Adjust the instructions

If a problem repeats across several questions, correct the instructions instead of adding isolated exceptions. Clearly define the role, allowed sources, restrictions, and escalation criteria.

Select Save changes before starting another test.

Review knowledge

When the agent answers a specific fact imprecisely:

• Confirm the document is indexed.
• Check that the knowledge base is selected.
• Look for conflicting or outdated documents.
• Improve titles and structure in the original document.

Start a new chat

Select New chat to remove temporary context and test from a clean state. Do this after significant instruction or knowledge changes.

One correct response does not mean the agent is ready. Repeat the test with different wording and edge cases.