What are the key metrics and measurement strategies that organisations should monitor to ensure their AI agents behave reliably, safely, and usefully?