How do I set up alerts for my automations without learning to ignore them?

Automations fail in two annoying ways. They go quiet and you never find out, or they shout so often you stop listening. I've had both.

What happens when nothing alerts you?

You lose days. My call recorder wasn't working last week and I didn't know. It was just logged out, and it never told me. Infuriating.

The fix: a check every weekday that confirms the recordings are happening, so I never go three days without knowing it failed. And that alert has to stand out from the routine ones, because we get a lot of alerts and this one is worse than most.

Why do noisy alerts backfire?

Because you stop reading them. I got four or five bug emails about one of my sites in a short stretch while the site was up. If the site isn't down, I don't think I should be getting those.

"4 in an hour or two will make me ignore them, like i did"

How do you stop false alarms?

Go back to first principles. Either the system is working or it's not. If it's truly broken, I want to know so we can fix it. If it's working, I want to stop hearing that there may be problems, and I want the system to stop thinking there may be problems. Fix whatever keeps raising them instead of living with the noise.

How often should a background agent check in?

Less than you'd think. For an agent that looks for new videos I've saved, I set it to run three times a day: 9 am, 1 pm and 6 pm. That should make it better and cheaper. I don't need any reply if nothing is there.

The one exception: if more than 48 hours go by with nothing, tell me. Either the tool feeding it is broken, or I haven't saved anything and should post one just to make sure everything's good.

What should users see when something fails?

A plain-English reason, whenever it affects what they're doing. In my software, every error needs a human explanation shown to the user. A customer whose scraping didn't run should have seen something like: this isn't running because there are no funds in the scraping account, so we cannot scrape. A product hiding its failures is unacceptable, and I've told AI agents that at least ten times.

But don't show errors that don't change anything for the user. If one of 15 feeds isn't pulling right now and the rest are full of content, a warning at the top of the screen makes people assume there's a huge problem. Show the error when they actually try to use that source and it fails.

Which alerts are worth the interruption?

The ones where work stops. I got multiple notifications that an AI account was out of money, so things didn't run. That's exactly what I don't want, and the key projects can't be stopped.

Spending alerts cut the other way too. I keep automatic top-ups small, about $10, so if a key ever gets exposed I don't get hit with hundreds of dollars of charges in the middle of the night. My cap is around $300 for the entire system, with smaller limits on each key.