I have 3 articles that discuss a fairly inclusive methodology to doing this (I wanted to attache them, but it exceeds the maximum file size). They are also supposed to be posted in the download area, but it hasn't been completed yet.
Please email me directly and I'll send them to you.
An answer to your question is that both models you discuss have validity. The key is that no matter HOW you come up with your model, that the system/application should not be able to tell the difference between your modeled load during a test execution and the actual users in production. Idealistic, yes, but if you keep that in mind it will help you end up with a good model.
The important thing for me is to define business transaction rates, and check that they are valid and that you are recreating them.
e.g. - how many documents are modified in an hour ? by how many different people ? how long does each person spend doing that ?
Once you have answered those questions, you can define the number of virtual users required.
As far as multiple actions by single user compared with single actions by multiple users - I prefer the latter as it's easier to adjust the ratios in the scenario.