How to make anonymity work better

This amounts to having each agent choose whether to randomly forward a message to one of its peers, or deliver it to its destination. Presumably, the actual message contents (including its original sender) are encrypted in a way that only the receiver can read, and not the intermediate hops. This helps to ensure that an IP-address attack cannot succeed, and is similar to the techniques used by the Cypherpunk remailers.

Since we are implicitly assuming that whatever the agents are communicating is personal enough to protect with good anonymity and/or cryptography,, it is convenient that random reforwarding can also aid in foiling a perennial problem of communication in the face of malevolence, namely traffic analysis.

In particular, if a given user's agent has a large number of potential "first-hop" neighbors, and if most of those can be presumed at least moderately trustworthy (e.g, only a statistically small number of them are motivated to backtrace the connection by IP address), the actual amount of traffic data leaked is relatively small. [A policy of communicating random data to peer agents, to be thrown away after decryption in most cases, would be even greater protection, but is presumably not enough of a threat for most users---though conceivably this could be configurable by how willing individual users are in tolerating the overhead.]