Ditch the Boss: AI Agents Coordinate Directly

A new framework from Stanford, called a decentralized language model (DeLM), challenges the assumption that AI agents require a central orchestrator. DeLM enables agents to communicate and coordinate directly, bypassing the need for a central controller to route requests and merge progress. This approach aims to reduce inference costs and coordination latency.

The traditional multi-agent system relies on a main agent to break down tasks, distribute them, and synthesize results, which can create a communication bottleneck and potentially distort information. DeLM, conversely, uses a shared knowledge base for agents to build upon verified findings. Agents write compact, verified updates into this shared context, creating a shared problem state.

DeLM's process involves initializing tasks, parallel execution by agents, and compression and verification of results into reusable 'gists.' Only fully verified findings are shared. This framework has demonstrated significant improvements in real-world benchmarks, achieving 10.5% better performance on SWE-bench Verified and reducing cost per task by roughly 50%. It also showed the highest accuracy on LongBench-v2 Multi-Doc QA across multiple model families.