Having the intended behavior as an emergent property of a lot of independent actors is always better than programming it top-down in a way that can be controlled by a single entity or shutdown.
Unless you're trying to make a specific thing happen quickly. Emergent properties are are very hard to construct intentionally and are often unexpected. It's easier to reason about single-threaded programs or about multiple threads acting independently. Once you're talking about multiple threads of execution that don't act independently systems can get pretty hairy. I think that's the reason client-server/master-slave architectures dominate computing today. Here's hoping for a p2p future!
