Stanford's DeLM cuts multi-agent task costs 50% — without a central orchestrator — Blankdot