Why conditioning on "the agent takes action a" isn't enough — LessWrong