You can use "the UNIX way" and have one process per plugin. Simply create two FIFOs per plugin and have a long-running process connected to those two FIFOs.
This is turning out to be much more complicated than I had originally expected. With the UNIX way saying of "using the right tool for each job", I kind of wish there was a tool to build scalable network programs... Oh well, until that comes along, I'm going to stick to Bash.
He's using chipper sarcasm to illustrate the point that rewriting things "the Unix way" may produce a great troll fork on Github, but more up-front engineering is absolutely justified if you anticipate the need to scale.
To get more to the point, Unix pipelines in shell scripts are nifty but absolutely not designed to scale in the ways that a typical networked server process does.
I happen to agree with him; if the only point of hubot was to be a campfire bot with a plugin architecture, you could probably accomplish that in 5 lines of Ruby that would never see the light of day in a production environment.