An agent is never finished on day one. This article gives you a routine for testing before publishing and improving after launch.
The Test Agent section at the bottom of the agent editor is a private chat with the current draft. Nothing is sent to a customer and no AI Actions are charged, within the Free Test Limits shown in the panel (per user per day, per agent per day and per month).
Use Reset to start a fresh test conversation. Use Test Retrieval in Knowledge Sources to see which chunks a question finds; if the right chunk is missing, fix the source, not the prompt.
Screenshot: the Test Agent panel with a test question and the agent's reply
Every save creates a version. Under Version History you can compare previews and Restore an earlier prompt if a change made the agent worse. Publish only when the test chat looks right.
For the first two weeks, read the agent's conversations in the Team Inbox every day. Look for:
Most fixes are additions to the knowledge base: a service you forgot, a price that changed, a question nobody expected.
The AI Quality page in the AI section summarises Readiness, items that Need Attention, and suggested Actions for your account. Domain Scorecards show coverage and weak spots by business domain (pricing, booking, policies), including cases with bad feedback. Start with the lowest-scoring domain.
Use the AI analytics view to watch resolution rate, handoff rate and AI Actions used per answer over time. A rising handoff rate after a change usually means a knowledge gap.
Exact options may differ by plan.