Pitfalls of the agent model — LessWrong