Parallel agents are most useful where correctness matters least

Tasks that parallelize cleanly (docs, tests) are the ones where subtle quality matters least.