Copilot agents can take actions. But who verifies the results?

As Copilot agents become more autonomous, validation seems just as important as automation. How are you handling it?