Exploring Mild Behaviour in Embedded Agents — LessWrong