Why agents are powerful — LessWrong