Link actions and knowledge explicitly
Use
@Action: Name and @Knowledge: Name in your Specialist instructions. This makes the AI more reliable than just describing them in prose.Be specific with handoff conditions
“Customer is frustrated” is better than “when appropriate”. The more specific, the fewer surprises.
Test with the Simulator
The Simulator works for both email and chat. Use it to test a full conversation flow before making changes live.
Choose the right website lookup
Use website import for stable content like FAQs. Use live website lookup when you need dynamic URLs with parameters.
Use access levels correctly
“No sharing allowed” lets the AI use info for reasoning but not quote it. “Internal use only” hides it from the AI Agent entirely.
Require human approval for sensitive actions
Enable this on actions like refunds. The AI will still prepare everything, but a human confirms before execution.