Scaling Large-Scale Agent Networks: A Review of Topology, Memory, and Updates

Large language model (LLM)‑driven multi‑agent systems are expanding from software engineering to scientific analysis, web automation, organizational collaboration, and social simulation. As the number of agents and interaction complexity increase, some architectures sustain long‑chain, multi‑step tasks while others become unstable or inefficient.

Motivation

Agent marketplaces and deployed systems have grown from a few roles to dozens or hundreds of agents, indicating that large‑scale agent networks are moving from laboratory demos to open, persistent, real‑world environments. Data from Internet Archive snapshots of OpenAI GPTs, AWS Marketplace, and Agent.ai (Figure 1) show a rapid rise in both the number of agents available in marketplaces and the number of agents involved in a single task chain.

Three‑Dimensional Taxonomy

The paper introduces a three‑dimensional classification framework that characterizes large‑scale agent networks along:

Topology: centralized vs. decentralized .

Memory scope: global memory vs. local memory .

Update behavior: static vs. dynamic .

Combining the three axes yields eight typical categories (Figure 2). The authors observe:

Centralized systems simplify scheduling and consistency but risk a central bottleneck as scale grows.

Decentralized systems enable emergence and flexibility but can suffer local miscoordination and information drift.

Global memory provides shared context and state alignment; local memory mirrors distributed environments but may produce divergent views.

Static systems are easier to analyze and reproduce; dynamic systems better support long‑horizon tasks and adaptive collaboration.

World‑Model Inconsistency

Beyond communication protocols, the deeper bottleneck is inconsistency among agents' world models. Even with perfect message transmission, agents may interpret the same statement or state differently because of divergent internal knowledge, preferences, or memories. This misalignment propagates through the system, causing:

Belief drift at the cognitive level.

Unstable cooperation at the behavioral level.

Goal divergence at the task level.

Non‑stationary dynamics that hinder convergence at the system level.

Future Research Directions

The authors highlight four research avenues derived from the taxonomy and world‑model analysis:

Develop clearer consistency models.

Strengthen shared‑state control mechanisms.

Design more mature routing and communication‑scheduling strategies.

Address identity, security, and robustness for open environments.

They also note that current evaluation suites are limited to small‑scale benchmarks, whereas future real‑world systems may involve thousands to millions of agents.

Conclusion

The taxonomy demonstrates that scalability depends not merely on the number of agents but on the interplay of topology, memory scope, update dynamics, and especially alignment of world models. Understanding these dimensions provides a structural map for designing and evaluating large‑scale multi‑agent systems.

Paper links: https://www.techrxiv.org/doi/full/10.36227/techrxiv.177127384.46731320/v1 ; https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6390059

Code example

来源：PaperWeekly
本文
约3000字
，建议阅读
5
分钟
多智能体越来越强？