A typical Rube Goldberg machine is scalable to build because only specific pairs of components interact. You build A and B and hook them up. Then you build C and hook it up to B, and so on. You don't have to worry about how to connect A and C. So building a machine with N components requires that you get Θ(N) interactions to work.
Unfortunately, the same is not true for reliability, which is why Rube Goldberg machines are "hard". If each step works 99% of the time, then 10 components (+ 9 interactions) work together only (0.99)^19 = 83% of the time. 20 components puts you at 67% and with 40 your machine will fail more than half the time even if you got everything nearly perfect!
Software engineering would love to be doing that well. Each software component may interact with many other subsystems, forming a general graph rather than a chain or tree. Fortunately, computers tend to be highly deterministic environments rather than random ones, and software components can compensate for bad behavior elsewhere in the system. Otherwise nothing would ever work at all.