Is it possible to train an agent to take and pass a multiple-choice exam based on a digital version of a textbook for some area of study or curriculum? What would be involved in implementing this and how long would it take, for someone familiar with deep learning?
There are programs that do this today, for some values of "curriculum" and "exam". It does not even require deep learning; a simpler information retrieval algorithm and some rules for composition work and achieve high scores on machine graded essays.
For human graders, there is research on automatically generating essay-length text responses to queries in a certain domain.
Both linked applications are rule-based rather than based in deep-learning. I'd guess that a deep-learning approach would be much less efficient (in computer resources) in producing comparable results.
I guess it could be possible with a lot of questions to learn from and only from a certain topic. Just watch what IBM was capable of https://www.youtube.com/watch?v=WFR3lOm_xhE very impressive!