AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems

AgentNet is an arXiv paper (submitted 1 Apr 2025, revised 29 May 2025) proposing a decentralized alternative to the central-orchestrator pattern in LLM-based multi-agent systems. Instead of a controller routing work to static roles, agents self-organize into a dynamically structured Directed Acyclic Graph (DAG): each agent adjusts its own connectivity and routes tasks based on local expertise and current context.

Three claimed innovations: (1) fully decentralized coordination with no central orchestrator, aimed at robustness and "emergent intelligence"; (2) a real-time adaptive graph topology so the collaboration structure reshapes itself per task rather than being fixed; (3) a retrieval-based (RAG) memory per agent that supports continual skill refinement and specialization over time.

The motivating argument is familiar from systems engineering: centralized coordination creates scalability bottlenecks, single points of failure, and reduced adaptability. Privacy and proprietary-knowledge concerns additionally block cross-organizational collaboration, producing siloed expertise — AgentNet's minimal data exchange is pitched as enabling fault-tolerant, privacy-preserving collaboration across organizational boundaries.

Experiments reportedly show higher task accuracy than both single-agent and centralized multi-agent baselines. Note: the ingested source is the arXiv abstract page only, so all claims here are the authors' own framing, not evaluated.