At what point do you stop trusting your automation scripts with database writes?

It’s fun setting up AI agents to handle routine backend tasks until you realize how easily a weird input can mess things up.

Even with good guardrails giving an agent direct write access to a live database always feels like playing with fire. I've started forcing a manual approval step for anything that modifies or deletes existing records no matter how confident the script claims it is.

Where do you personally draw the line do you let agents touch production data automatically or do you keep a human check on every single write?