Operate
Test the agent, send business events, hand threads between the agent and your inbox, and run evaluations.
Operate is the working page: everything here exercises a live agent without necessarily involving a customer.
Test the Agent
Test message sends a message to the agent and shows the reply. No real customer is notified, so it is the safe way to check a new skill or connector tool.
Send a Business Event
A business event tells the agent something happened on your side, so it can act on it in an ongoing conversation.
| Field | What goes in it |
|---|---|
| Event type | A stable key you choose, reused for every event of this kind |
| Consumer phone number | E.164 format, including the country code |
| Description | What this event means, in words the agent can act on — for example "The customer's order has left the warehouse." |
| Payload | The structured data that goes with it |
Thread Control
Hands a live conversation between the AI agent and your inbox. Use it to pull a conversation to a human mid-thread, or hand one back once the human is done.
Run an Evaluation
Meta simulates conversations against your scenarios and scores how the agent handled them.
Load eval cases fetches the scenarios; Reload cases refreshes the list; a run reports as QUEUED, RUNNING, then a result or FAILED.
Evaluations are how you tell whether a change to knowledge or skills actually improved anything. Run one before and after a substantial edit rather than judging by a single test message.
Related Pages
- Agent Settings — rollout and audience
- Connectors — the log panel, when a tool call fails
- Inbox — where a handed-off conversation is answered