They ride shotgun to the main app, checking in on it, and taking action when things go south. They know how to check in and how to make sure that things are going okay. They know how to call for help if things go south. Maybe they call in to pagerduty for you automatically after N failures over a period of time. Maybe they just mark the instance unhealthy. That's up to you.
Sidekicks could be configured via standard env vars, being told if alerts are warranted on failure, and if they should page, whom deserves it.
Ideally, you have very few of these, as your apps conform to a common convention.
This aims to eliminate the need for a centralized monitoring system, to make things more composable, and still give you the benefits you are looking for.