1. LLM-Based MAS Taxonomy

Based on Multi-Agent Collaboration Mechanisms: A Survey of LLMs

https://mp.weixin.qq.com/s/GGZ3FuHn2Fcv7U3xW3G3Qg

Dimension Categories Example Systems
Actors Task Executors, Planners, Evaluators, Aggregators AutoGen (Planner & Executor), AgentVerse (Evaluator)
Collaboration Types Cooperation, Competition, Coopetition MetaGPT (Cooperation), LLM Debate (Competition), MoE (Coopetition)
Structures Centralized, Decentralized, Hierarchical AutoGen (Centralized), Debate Models (Decentralized), CAMEL (Hierarchical)
Strategies Rule-Based, Role-Based, Model-Based CAMEL (Rule-Based), MetaGPT (Role-Based), ToM AI (Model-Based)
Coordination Static, Dynamic Sequential Agents (Static), AutoGen Adaptive Roles (Dynamic)

2. Coordination Evolution

Method Key Technologies Advantages Limitations Example Systems
Rule-Based Coordination Predefined rules, scripts Simple, predictable Rigid, non-adaptive CAMEL, AutoGen
Evolutionary Search Genetic algorithms, ES, NAS Adaptive, no manual tuning High computation EvoAgent, AutoAgents
Reinforcement Learning (RL) MARL, HRL Learns optimal coordination Long training time GPTSwarm, AutoML
Agentic Supernet (MaAS) Probabilistic optimization Adapts dynamically, reduces cost New, needs research MaAS framework
LLM-Driven MAS Large Language Models, NLP Natural coordination, human-like reasoning Hallucination, memory issues MetaGPT, AgentVerse

1. Traditional Rule-Based Coordination

e.g: CAMEL, AutoGen

Advantages: Simple to implement, predictable behavior.

Limitations: Not adaptive to new environments, rigid structures.

2. Evolutionary and Search-Based Methods