Before delving into a detailed discussion of how mailers are expected to do mail routing, it would seem to make sense to give a brief overview of how this memo is approaching the problems that routing poses.
The first major principle is derived from the definition of the preference field in MX records, and is intended to prevent mail looping. If the mailer is on a host which is listed as an MX for the destination host, the mailer may only deliver to an MX which has a lower preference count than its own host.
It is also possible to cause mail looping because routing information is out of date or incomplete. Out of date information is only a problem when domain tables are changed. The changes will not be known to all affected hosts until their resolver caches time out. There is no way to ensure that this will not happen short of requiring mailers and their resolvers to always send their queries to an authoritative server, and never use data stored in a cache. This is an impractical solution, since eliminating resolver caching would make mailing inordinately expensive. What is more, the out-of-date RR problem should not happen if, when a domain table is changed, affected hosts (those in the list of MXs) have their resolver caches flushed. In other words, given proper precautions, mail looping as a result of domain information should be avoidable, without requiring mailers to query authoritative servers. (The appropriate precaution is to check with a host's administrator before adding that host to a list of MXs).
The incomplete data problem also requires some care when handling domain queries. If the answer section of a query is incomplete critical MX RRs may be left out. This may result in mail looping, or in a message being mistakenly labelled undeliverable. As a result, mailers may only accept responses from the domain system which have complete answer sections. Note that this entire problem can be avoided by only using virtual circuits for queries, but since this situation is likely to be very rare and datagrams are the preferred way to interact with the domain system, implementors should probably just ensure that their mailer will repeat a query with virtual circuits should the truncation bit ever be set.