Stanford's DeLM cuts multi-agent costs 50% by killing the central orchestrator

Stanford researchers Yuzhen Mao and Azalia Mirhoseini released DeLM, a decentralized multi-agent framework where sub-agents coordinate directly through a shared knowledge base instead of routing every interaction through a central controller. The architecture cuts coordination cost roughly 50% on benchmark tasks while reducing the information loss that comes from a main agent merging and rebroadcasting partial findings. The authors argue the conventional "boss-agent" pattern scales poorly because each subtask report becomes a serialization bottleneck and the controller often dilutes or distorts evidence. DeLM lets agents "build on prior findings, avoid repeated failures, preserve constraints, and recover detailed evidence only when needed."

View full digest for June 17, 2026