Protocols and multimethods are great tools to manage polymorphism, whereas Graph is about composition. We use both extensively in our codebase, and treat them as separate tools in our toolbox for building fine-grained, composable abstractions.

For example, I don't think protocols or multimethods could easily do any of the things mentioned in the second half of the post (execute part of a computation, auto-parallelize it, monitor the components, etc).

That said, there is actually one case where we use Graphs to solve a difficult polymorphism problem, which I discussed a bit in my Strange Loop talk. Our core newsfeed generation logic used to be composed of protocols/multimethods (we tried both), since each feed type (we have about 10) can define different variants of various steps in the pipeline (but most of the steps are the same). This worked fairly well, but as our system grew more and more complex, we found that there was still a lot of overhead, since the protocol had to contain all the steps that could change, leading to lots of extra complexity.

We've replaced all of this with Graph, where we just define an 'abstract' graph with the most common steps, and each feed type modifies the graph by changing or adding steps -- and we've found this way to be much simpler and easy to understand than what we had before.

This case is special, since it involves both a complex composition and polymorphism. Everywhere else in our codebase, we use (and love) protocols and multimethods for polymorphism.