Language Agents Reduce the Risk of Existential Catastrophe — LessWrong