Even that approach to networking is, dated.
Mesh networking has a higher level of required complexity per unit but does not require anything like central nodes, just enough nearby units to allow a path to the required system or data that the local unit requires to complete a task.
If I was designing a thing that was capable of being autonomous, yet being part of a whole, I would put in all required systems but in such a way that they could work on sync with neighboring units.
Add in a communication system that has zero delay [already demonstrated at the lab scale] and you get something that is truly terrifying in potential.
How small a package and what form should it take?
Well that depends on the systems incorporated and their inherent bulk.
Start with a positronic brain, add communications, sensors, tractor beam, deflector system, replicator, warp drive, gravity drive [how else do you get the damn things to float around in an atmosphere?]
I am thinking a cube is impractical, what with the issues of edges and corners, and a sphere as an inherently stable geometry.
How big? Now that is a good questions. Just how small could one reduce such systems and still have them be functional on a practical level?
I remember one star trek episode where it was pointed out that the not quite dead yet aliens with their tech could make an enterprise class warp drive the size of a walnut.
If the tech were that advanced I could see them as the size of a softball overall.
One unit would be the equivalent of a runabout or shuttle in capability. Cluster a few hundred and you get a cruiser. A cluster of a few million and you get a battle station.
.... Hmm, most of those systems would facilitate interactive holographic projection.
Now that could allow for some very interesting anthropological research and interaction with all sorts of intelligent species.
Or just hide as innocuous gas giants throughout the Galaxy and stay well away from those crazy bags of mostly water.