Tools and reporting for an AI phone agent: booking, a safe send gateway and call summaries

The booking scenario is a webhook with an action field: slots, book, reschedule or cancel. For slots it checks the calendar is active, reads the next fourteen days of free time, and a code step keeps only weekday working hours before answering. Every action ends in its own webhook response, so the agent always gets a definite answer.

Before booking, it checks again that the slot is still free and that the request matches a slot it offered; an unknown customer gets a hold answer, not a booking. The appointment is created with no retry, so a slow response cannot make two, then read back to prove it exists before the deal moves to the estimate-scheduled stage and the agent is told it is booked. Rescheduling and cancelling read the appointment first and read it back after the change.

Every outbound message goes through one send scenario. It opens with a kill switch and transport defaults in a set-variables step, reads the contact's tags and do-not-disturb status, and a code step chooses the transport: stub, unless every guard passes. The stub route logs what would have been sent to a data store and adds a note to the contact, so a dry run leaves a visible trace.

A QA route sends real texts only to an allow-listed test contact. A contact who has stopped messages or is on do-not-disturb, or a message with no text, is skipped. Each route answers the caller with its own status (stubbed, sent to QA, sent or skipped), so the agent's log says exactly what happened.

After each call the voice platform posts to a webhook; one scenario filters out calls that need no summary, iterates the recipients and emails a structured summary, with a retry handler if the mail server refuses. A second sends the platform's daily report and review digest by email. For the second agent, a router posts call records and alerts to a messaging channel and compiles a daily report.