How do you ensure human oversight in AI agent actions?

In the realm of AI development, particularly when building agents that can perform critical tasks autonomously, ensuring human oversight is vital. A recent project called Approv addresses this by implementing a human approval step for risky actions taken by AI agents. It uses a simple API call to pause actions and prompts a human for approval via WhatsApp, ensuring a tamper-evident audit trail. What strategies do you use to balance automation with necessary human intervention in your AI projects?