Here's one very concrete example of something I like:
Their nodes have the Pd / Max / vvvv / etc style of a row of inputs on the top, and a row of outputs on the bottom. In Nodes, they've devoted the entire node box graphic to inputs and outputs, by putting the node name off to the side. Each input and output is a square. There's no padding. This means the hit target for each input and output is as large as can be given the size of the node and the uniformity of the squares. That makes patching less fiddly than it is in tools with smaller hit targets.
Here's a criticism:
All their edges appear to be S-curves. That's fine as a default, but it definitely leaves a huge part of the possibility space unexplored.
When evaluating these sorts of VPLs, the devil is in those sorts of details.