Steve Keen takes up the cudgels. (Via.)
There is a pattern to neoclassical attempts to increase the realism of their models... The author takes the core model – which cannot generate the real world phenomenon under discussion – and then adds some twist to the basic assumptions which, hey presto, generate the phenomenon in some highly stylised way. The mathematics (or geometry) of the twist is explicated, policy conclusions (if any) are then drawn, and the paper ends.
The flaw with this game is the very starting point, and since Minsky put it best, I’ll use his words to explain it: “Can 'It' – a Great Depression – happen again? And if 'It' can happen, why didn't 'It' occur in the years since World War II? ... To answer these questions it is necessary to have an economic theory which makes great depressions one of the possible states in which our type of capitalist economy can find itself."
The flaw in the neoclassical game is that it never achieves Minsky’s final objective, because the “twists” that the author adds to the basic assumptions of the neoclassical model are never incorporated into its core. The basic theory therefore remains one in which the key phenomenon under investigation ... cannot happen. The core theory remains unaltered – rather like a dog that learns how to walk on its hind legs, but which then reverts to four legged locomotion when the performance is over.Right.
Any theory is an abstraction of the real world, but the question is which features of the world you can abstract from, and which, for the purposes of theory, are fundamental. Today's consensus macroeconomics  treats intertemporal maximization of a utility function (with consumption and labor as the only arguments) under given endowments and production functions, and unique, stable market-clearing equilibria as the essential features that any acceptable theory has to start from. It treats firms (profit-maximizing or otherwise), money, credit, uncertainty, the existence of classes, and technological change as non-essential features that need to be derived from intertemporal maximization by households, can be safely ignored, or at best added in an ad hoc way. And change is treated in terms of comparative statics rather than dynamic processes or historical evolution.
Now people will say, But can't you make the arguments you want to within standard techniques? And in that case, shouldn't you? Even if it's not strictly necessary, isn't it wise to show your story is compatible with the consensus approach, since that way you'll be more likely to convince other economists, have more political influence, etc.?
If you're a super smart micro guy (as are the two friends I've recently had this conversation with) then there's probably a lot of truth to this. The type of work you do if you genuinely want to understand a labor market question, say, and the type of work you do if you want to win an argument within the economics profession about labor markets, may not be exactly the same, but they're reasonably compatible. Maybe the main difference is that you need fancier econometrics to convince economists than to learn about the world?
But if you're doing macroeconomics, the concessions you have to make to make your arguments acceptable are more costly. When you try to force Minsky into a DSGE box, as Krugman does; or when half of your paper on real exchange rates is taken up with models of utility maximization by households; then you're not just wasting an enormous amount of time and brainpower. You're arguing against everyone else trying top do realistic work on other questions, including yourself on other occasions. And you're ensuring that your arguments will have a one-off, ad hoc quality, instead of being developed in a systematic way.
(Not to mention that the consensus view isn't even coherent on its own terms. Microfoundations are a fraud, since the representative household can't be derived from a more general model of utility maximizing agents; and it seems clear that intertemporal maximization and comparative statics are logically incompatible.)
If we want to get here, we shouldn't start from there. We need an economics whose starting points are production for profit by firms employing wage labor, under uncertainty, in a monetary economy, that evolves in historical terms. That's what Marx, Keynes and Schumpeter in their different ways were all doing. They, and their students, have given us a lot to build on. But to do so, we  have to give up on trying to incorporate their insights piecemeal into the consensus framework, and cultivate a separate space to develop a different kind of economics, one that starts from premises corresponding to the fundamental features of modern capitalist economies.
 I've decided to stop using "mainstream" in favor of "consensus", largely because the latter term is used by people to describe themselves.
 By "we," again, I mean heterodox macroeconomists specifically. I'm not sure how much other economists face the same sharp tradeoff between winning particular debates within the economics profession and building an economics that gives us genuine, useful knowledge about the world.