Well, you can construct the following protocol for fully untraceable communication between n nodes, numbered i \in 1...n: Divide time into intervals of size t, where t is larger than the time it takes to propagate a message of constant size s from any node to all of the remaining n-1 (e.g. by flooding). For interval j, node i = j % n always transmits a message m of size s to all other nodes with the following characteristic: m is either the output of a PRG or a message encrypted with the key for a host k to which i wishes to communicate a message, the choice of which is entirely up to i. Under this scheme - assuming previously set up authentication between every pair of nodes and an encryption scheme in which encrypted messages are indistinguishable from random data without the decryption key - any node can send a message to any other node in such a way that no one else inside or outside this network can know the contents of the message or even that the communication took place. For any node not receiving communication, the protocol would be indistinguishable of one in which all transmissions are random noise.
Of course, the issue is that latency in this scheme is O(n) and per-node bandwidth is O(1/n), with large constants. Also, it's a reasonable suspicion in practice that no one would set up this scheme and then actually have zero communication going on over it, so it still reveals that "at least one of the n nodes is talking to at least another of the n nodes".
Or use multiple layers so at least you are less likely to be targeted by larger more automated operations, and only so much at risk to a targeted attacked.
