Add "You should split each connection over many paths"
The answer should mention Network Team new research about traffic congestion.
We don't currently think this is a good idea. You see, the attacks we're worried about are at the endpoints: the adversary watches Alice (or the first hop in the path) and Bob (or the last hop in the path) and learns that they are communicating.
If we make the assumption that timing attacks work well on even a few packets end-to-end, then having more possible ways for the adversary to observe the connection seems to hurt anonymity, not help it.
Now, it's possible that we could make ourselves more resistant to end-to-end attacks with a little bit of padding and by making each circuit send and receive a fixed number of cells. This approach is more well-understood in the context of high-latency systems. See e.g. Message Splitting Against the Partial Adversary by Andrei Serjantov and Steven J. Murdoch.
But since we don't currently understand what network and padding parameters, if any, could provide increased end-to-end security, our current strategy is to minimize the number of places that the adversary could possibly see.