Towards a formalization of the agent structure problem — LessWrong