Skip to content

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

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AgentNet is a fully decentralized multi-agent framework where LLM agents self-organize via a dynamically evolving Directed Acyclic Graph (DAG), routing tasks peer-to-peer without any central orchestrator. Each agent maintains separate RAG-based memory pools for routing and execution, enabling continual specialization from task trajectories. The system outperforms centralized multi-agent baselines (MetaGPT, AFLOW, GPTSwarm) on math, logic, and API-calling benchmarks across three backbone LLMs.

Problem

Existing LLM multi-agent systems depend on a central orchestrator with static, predefined workflows and fixed agent roles. This introduces a single point of failure, limits scalability, prevents dynamic role adaptation, and makes cross-organizational collaboration infeasible due to privacy and data-ownership concerns (each institution cannot expose proprietary knowledge to a shared controller).

Method

AgentNet is defined as a graph G = (A, E) where agents dynamically adjust connectivity via a weight matrix updated after each task: w_{m+1}(i,j) = α·w_m(i,j) + (1−α)·S(a_i, a_j, t_{m+1}). Edges below a threshold θ_w are pruned, causing the topology to evolve toward efficient collaboration paths.

Each agent has two modules—a router (routed via RAG over routing-trajectory memory M^rou) and an executor (augmented via M^exe)—both backed by the same LLM using the ReAct framework. Memory fragments f = (observation, context, action) are retrieved by semantic embedding similarity. When a memory pool hits capacity C_max, an LLM-driven pruning step removes low-utility trajectories.

Task allocation selects an entry agent by matching task capability requirements (Φ(o_t)) against agent capability vectors (cv_i), updated via exponential moving average after each task. At each hop, the router chooses one of three operations: Forward (pass unchanged to a better-suited agent), Split (decompose and delegate subtasks while executing matching portions), or Execute (complete locally). The DAG constraint prevents routing cycles.

Key Contributions

  • Fully decentralized coordination with no central orchestrator; routing authority is distributed across all agents.
  • Dynamic DAG topology: edge weights evolve from task co-success, with periodic pruning to remove unproductive connections.
  • Dual RAG memory per agent (separate pools for routing vs. execution), supporting continual specialization without explicit role assignment.
  • Privacy-by-design: only minimal task-relevant metadata passes between agents; sensitive trajectories remain local.
  • Empirical demonstration that decentralized evolutionary coordination outperforms both single-agent and centralized multi-agent baselines.

Results

  • MATH (DeepSeek-V3, 3 agents): AgentNet 92.86% vs. MetaGPT 92.14%, AFLOW 91.67%, GPTSwarm 72.14%.
  • BBH (DeepSeek-V3, 3 agents): AgentNet 94.00% vs. GPTSwarm 90.00%, AFLOW 88.00%, MetaGPT 64.00%.
  • API-Bank (DeepSeek-V3, 3 agents): AgentNet 30.00% vs. AFLOW 28.00%, MetaGPT 22.00%.
  • Qwen-turbo / BBH: AgentNet 92.00% vs. GPTSwarm 75.00%, AFLOW 57.00%.
  • Ablation — evolution phase (GPT-4o-mini): removing evolution drops MATH from 85.00 to 77.86, BBH from 86% to 76%, API-Bank from 32 to 23.
  • Heterogeneity (BBH): with 5 agents, heterogeneous configurations (skill/LLM/both) outperform fully homogeneous (0.86 vs. 0.79–0.81 acc); with 3 agents, homogeneous is best (0.86).
  • Scalability: BBH test accuracy peaks at 86 with 9 agents / 40-executor pool, rising from 80 at 3 agents / 30 executors.

Limitations

  • Performance gains from adding more agents show diminishing returns, suggesting an optimal team size beyond which overhead offsets benefits.
  • With small teams (3 agents), heterogeneous configurations underperform homogeneous ones, limiting the diversity benefit to larger deployments.
  • The DAG routing constraint prevents cycles but also restricts certain collaborative patterns that might require iterative agent interaction.
  • Experiments use at most 9 agents; scalability claims for very large networks (dozens–hundreds of agents) are not empirically validated.
  • No ablation on the RAG retrieval quality or embedding choice, leaving unclear how sensitive the system is to embedding model selection.
  • Privacy guarantees are architectural/design arguments, not formally proven; no adversarial privacy evaluation is conducted.

Relevance to Agentic AI / LLM Agents

AgentNet directly advances the architecture of multi-agent LLM systems by replacing the dominant centralized-orchestrator paradigm with fully peer-to-peer coordination, addressing a known scaling and fault-tolerance gap in frameworks like AutoGen and MetaGPT. The dual RAG memory (separate pools for routing vs. task execution) is a concrete mechanism for enabling emergent specialization without pre-assigned roles, which is a key open question in agentic system design. The dynamic DAG topology and capability-vector-based task routing provide a principled alternative to static workflow graphs, connecting to broader work on adaptive and self-organizing agent networks (GPTSwarm, EvoMAC). For researchers tracking agentic AI, this paper provides both a deployable framework and ablation evidence for the value of decentralized routing and continual memory in complex, multi-step task settings.