What the heck do I alert on anymore?

My AI agents make thousands of decisions I can't supervise. Here's how I'm figuring out what suspicious behavior looks like and what to alert on.