The students would learn a lot though. Most CS courses will give theory and practice for algorithms etc, but only give theory for software design. Trying to teach a student how and why it's important to encapsulate logic is generally a losing battle until you put them in a position where it makes sense. It's just a theoretical topic (that sounds like a religion) to not have global state until you dump the student in with 100 others all declaring global state. Same for namespaces and interfaces.

It would be a really interesting experiment even: divide up a big problem into 100 or so features and assign each feature to a random person. Give the students some time to work out who they need their code to communicate with and draw up some simple interfaces. Use github or similar with an account for every student. Make the TAs in charge of accepting pull requests. It could be the main focus of a software engineering course.

EDIT: The students wouldn't necessarily need to work with all 100 others - only the ones whose feature their feature interacts with. eg, 1 student could need to receive input from 5 inputs and calculate some statistics - he would only need to talk to the 5 students who would feed him data, and then talk to the 1 student he would feed data to.

You could go halfway, and have the student's projects all expose / consume web APIs to one another.

Of course then you'd have the student that forgot to include SQL injection protections getting their project goat.se'd during a demonstration.