One major benefit of protocols (IMHO) is that the protocol methods are properly namespace qualified. Thus I can have multiple protocols in different namespaces that define a foo method and extend them all (or a subset of them) upon existing types. However, that's not true with extending protocols inline with defrecord and deftype, or with extending protocols on the Java side by implementing their interfaces.
I know that I get the error because both the Foo and the Bar interfaces backing the protocols have a mymethod method and thus they cannot be implemented both at once (at least not with different behavior).
But why throw away the namespacing advantages we have with protocols? E.g., why is the protocoltest.foo.Foo method not named protocoltest$foo$mymethod (or some other munged name) in the corresponding interface? That way, both methods can be implemented inline where you gain the speed advantage, and you can do the same even from the Java side. (Currently, invoking clojure.core.extend from the Java side using clojure.java.api is no fun because you have to construct maps, intern keywords, define functions, etc.)
Of course, the ship of changing the default method naming scheme has sailed long ago, but maybe a :ns-qualified-method-names option could be added to defprotocol.