Not to be all Winnebago Man here, but all the code in Urbit (protocols and language alike) can be replaced and enhanced, except Nock. In fact, the normal way to drive this process is just by merging the new version of the code. It's certainly true, however, that you don't want locked-down components.

Urbit is very happy to be an HTTP client and server. (Alas, HTTP has become the "everything protocol.") It also has its own network protocol, which it prefers for conversations between urbits.

The unreliability of network packet delivery (or more exactly, the unreasonable effectiveness of unreliable networking) is basically the greatest technical idea of the 20th century as far as I'm concerned.

There is actually a significant difference between processing unordered, unreliable packet streams, and ordered, reliable, message streams, with the "universal design." Neat thing about packets: you can drop them. If a packet sends the server into an infinite loop? Or at least, a timer detects an infinite loop? Drop the packet. From the network's perspective, the CPU is simply detecting congestion.

You can refuse to process a message, but that's not a stateless and passive operation; you have to return a failure report. But dropping a bad packet is the most wonderful thing in the world. By definition it doesn't change your server's state. It also doesn't leak information to the sender...