The protocol drawn on napkins
By the late 1980s the protocol that connected separate networks to the backbone, EGP, was showing its age. It could only describe a simple tree of networks, and the real topology was turning into something much less tidy.
In January 1989, at an IETF meeting in Austin, two engineers — Kirk Lougheed and Yakov Rekhter — sat down over lunch and sketched a replacement on a couple of napkins. The idea was to have routers exchange not just which networks they could reach, but the whole path of autonomous systems a route had travelled. That path made loops easy to detect and gave operators something to write policy against.
The "two-napkin protocol" was meant as a stopgap. It is still carrying the global routing table.
The first specification was published as an RFC a few months later. The protocol has been revised several times since, but the core design from that lunch — paths of autonomous systems, policy decided locally by each network — has survived every attempt to replace it.
The cost of being good enough
BGP was built for a few hundred cooperating networks that broadly trusted one another. It now connects tens of thousands that do not. Many of the large outages of the past decades come down to a router somewhere announcing a path it should not have, and everyone else believing it. Temporary designs that work tend to become permanent; this one is the best-known example.