The biggest risks include incorrect decisions, access to sensitive information, and unintended actions. 

To reduce these risks, AI agents should have clear permissions, human approval for important actions, activity logs, and regular monitoring to ensure they behave as expected.