What if the entire system is visible and editable live within a single IDE? - the "entire system" including everything that's running within clients of your service and everything that's running on all your scaled out servers .. and everything that's running in your "IDE".
Without having to make it smalltalk, can we borrow this idea today and implement it in JS?
In the short term, I can imagine a system which exposes a REPL from each backend instance and a websocket-driven REPL from each browser client, and ties them together in a unified interface that allows enumeration and provides convenient syntax for evaluation against multiple targets. Such a system might offer some interesting possibilities around realtime client and server hot-patching and problem investigation, if it could be made bulletproof enough not to also serve as a gigantic SPoF or security vulnerability. But I don't know if it's really similar to what you're thinking of.
I may be underestimating the complexity of this, but it does seem doable with some clarifications to concepts. For example, in the smalltalk world, every GUI object created is expected to be inspectable and modifiable. If we then require that we want to call a button instantiated in every browser an independent object, that would be a misfit. It might be adequate to treat that button presented in the IDE context as the object that manifests on all connected browsers. So when I look for messages sent by the button, I tap into the logstream for those events that the system received from all browsers connected. But when I manipulate its appearance, it reflects on all connected browsers. I don't literally need to tap into each connected browser client.