A practical observation: the economists who get most bent out of shape at the notion that maybe we don’t always have to derive everything from optimizing individual agents also tend, with remarkable regularity, to be the economists who make simple, ludicrous conceptual errors when they discuss real-world macroeconomic issues. See many posts here and on Brad DeLong’s blog for examples. I don’t think this is an accident; it really helps your ability to think clearly to have those simplified, ad hoc models always in the back of your mind.One of the challenges of models is that they need to make predictions that are true. If not, then they are (at best) descriptions of some previous time point. But if we want to inform policy, then it is essential that models are a good description of reality. So if the model makes a prediction that is obviously untrue then that should be a major red flag.
In my own view, the idea that people maximize "utility" as a decision making process is clearly wrong for any clear definition of utility (the idea that we maximize something that is an unobserved latent variable including social, financial, and cultural factors -- with arbitrary and time-varying weights -- is useful conceptually but unlikely to be a tractable variable for models). So building models with this as a foundation may or may not improve predictive power.
On the other hand, Newton's laws of physics were successful long before quantum and statistical mechanics were developed. So useful models may precede an understanding of the underlying microfoundations, at least in theory.
So, needless to say, I agree with Dr. Krugman.