[tomn]

I think Erlang has them for a more specific reason: in general it eschews all forms of data definition (records are just a fancy syntax for tuples with an atom at the start), which makes hot code reloading and transparent network communication simpler.

Both cases would require some synchronization of data structures (across time or over the network), and with user-defined types this can get complicated, and atoms make the lack of user-defined types much more pleasant (and more performant than strings).