Here it is:
There is no such thing as a purely micro-founded macro model that is actually useful for forecasting.
The original RBC models were, more or less, purely micro-founded models. But, they didn't track anything. They could match some unconditional moments, but couldn't replicate realistic dynamics.
People, we don't even have very good micro foundations for money! We just put it in the utility function or arbitrarily assume a "cash in advance" constraint. That's one reason why the original RBC models didn't even contain money.
Amazingly to me, Central Banks in the Western world have spent a lot of money and economist-hours trying to construct dynamic stochastic general equilibrium (DSGE) models that are actually useful for forecasting.
This effort has largely led to the de facto abandonment of micro-foundations. In the quest to make the models "work" we often either choose whatever micro-foundation that gives the best forecast regardless of micro evidence about whether or not it is accurate, or we just add ad hoc, non-micro-founded "frictions" to create more inertia. Or we just add more and more "shocks" to the model and say things like, "much of the variation in X is caused by shocks to the markup".
So macro has to live in this weird world where, to correctly evaluate policy changes and welfare, we need to use fully micro-founded models that we know do not track reality, and to track reality, we need to come perilously close to giving up on micro-foundations altogether.
No wonder we are so often in a bad mood!