I think the difficulty you name of that mostly refers to the human hours (and eventually annoyance) of constantly fixing the broken thing.
Nowdays you could likely just have a monitor that identifies elevated failure rates and dispatches an agent with a cheap model, say Luna or DeepSeek V4.1 Flash, to analyze the challenges and failures and automatically write, test and deploy the fix.
Maybe combine it with a higher level agent that all three or five invocations checks for repeated manual work the agent does and turns it into a general purpose fix.
Nowdays you could likely just have a monitor that identifies elevated failure rates and dispatches an agent with a cheap model, say Luna or DeepSeek V4.1 Flash, to analyze the challenges and failures and automatically write, test and deploy the fix. Maybe combine it with a higher level agent that all three or five invocations checks for repeated manual work the agent does and turns it into a general purpose fix.
Sounds a lot more viable today.