To be clear I didn't say "don't model" - models can have value but all models are wrong because the map is not the territory. That's a first order problem, any representation of something is not the thing itself and bears no requirement to continuously and accurately represent it.

Another first order problem is you must chose from all the data that is perceivable what is pertinent to model because you can't perceive and model it all (at least not in complex systems) - that 'is perceivable' and 'choosing what is pertinent' act as filters that rarely are questioned. How do you know that 'what is pertinent' hasn't changed if your only way to reason about the system is through the model itself? How do you know you aren't perceiving something more relevant? Designing models to be sensitive to what we mean and not what we've stated is a very hard problem. The state space of reality is far bigger than we can sense and store.

A second order problem lies in human propensity for low energy states, when given a model or a metric we'll champion it as the truth or the way because it is easier than facing the complexity - but complex systems are crafty they adapt continuously. For example, your boss wants 100% test coverage and a dashboard has been created to report it to them... Fine, we'll test the getters and setters, we'll modify the dashboard code to ignore certain files, we'll write pointless tests that just exercise code but make no assertions... etc. They likely won't check provided the dashboard keeps reporting what they want to see.

Another second order problem is in complex adaptive systems the act of measurement changes the behaviour of the system itself. We know this intuitively, pull out a camera and start recording some strangers who were going about their business.

As for "what should we do instead" - complexity is all about context dependence... so it depends? If you are in a complex adaptive system, get involved! You can't know it all - so have fun with it.