Stanford's DeLM lets AI agents coordinate without a central controller, cutting multi-agent inference costs 50% and beating ...