Me and MagicalTux were discussing this. The concept we came up with to provide anonymity was:

A_x wants to send to B_x

A_1 sends 40 BTC to M
A_2 sends 10 BTC to M
A_3 sends 30 BTC to M

Spread over 1 week M does:

M sends 10 BTC to B_1
M sends 3 BTC to B_2
M sends 5 BTC to B_1
M sends 6 BTC to B_2
M sends 10 BTC to B_3
M sends 15 BTC to B_1
M sends 5 BTC to B_1
M sends 22 BTC to B_3

- The sums the mixnet sends never total the sums received. Its +/- a random amount (biased slightly towards - as the fee).
- Plausible deniability by spreading the funds over time.
- A_x provides several addresses for the B_x destination. M cycles through this list randomly then restarts once it's finished.

The only method to reverse engineer payments using this mixnet is through statistical analysis. If the payments are broken down enough, and many different users channel their payments through the system then it's very hard to rationalise what happened from the block-chain.