Overview of Multi-Agent Systems
Introduction
Multi-Agent Systems (MAS) study how multiple autonomous agents interact, collaborate, or compete within a shared environment. From classical distributed artificial intelligence to LLM-driven multi-agent systems, this field is undergoing a paradigm shift.
From Classical MAS to LLM Multi-Agent Systems
Classical Multi-Agent Systems
Wooldridge & Jennings (1995) defined four key properties of agents:
- Autonomy: Capable of acting without direct external intervention
- Social ability: Able to interact with other agents
- Reactivity: Able to perceive and respond to environmental changes
- Pro-activeness: Able to take initiative to achieve goals
Classical MAS research (1990s-2010s) primarily focused on:
- Rule-based agent interaction protocols (FIPA-ACL)
- Distributed problem solving (Contract Net Protocol)
- Strategic interaction under game-theoretic frameworks
- Swarm intelligence (ant colony, particle swarm)
The New Paradigm of LLM Multi-Agent Systems
LLM-driven multi-agent systems bring fundamental changes:
| Dimension | Classical MAS | LLM Multi-Agent |
|---|---|---|
| Communication | Structured messages (FIPA-ACL) | Natural language dialogue |
| Reasoning | Rule engines / logic reasoning | Large language model reasoning |
| Adaptability | Requires pre-programming | Prompt as programming |
| Capability scope | Domain-specific | General-purpose |
| Complexity | Limited state space | Nearly unlimited behavior space |
Multi-Agent Interaction Patterns
graph TB
subgraph "Cooperation"
A1[Agent A] -->|Collaborate| A2[Agent B]
A2 -->|Collaborate| A1
A1 & A2 --> G1[Shared Goal]
end
subgraph "Competition"
B1[Agent A] -->|Oppose| B2[Agent B]
B2 -->|Oppose| B1
B1 --> G2A[Goal A]
B2 --> G2B[Goal B]
end
subgraph "Coopetition"
C1[Agent A] <-->|Cooperate+Compete| C2[Agent B]
C1 --> G3A[Partially Shared Goals]
C2 --> G3B[Partially Conflicting Goals]
end
Cooperative Patterns
Multiple agents work together to complete a task:
- Division of labor: Different agents handle different subtasks
- Complementary cooperation: Different agents possess different capabilities
- Redundant cooperation: Multiple agents do the same thing to improve reliability
Competitive Patterns
Agents have conflicting goals:
- Debate: Agents hold different viewpoints and reach better conclusions through debate
- Red-blue teaming: Attack agent vs. defense agent
- Auction/bidding: Agents compete for limited resources
Coopetition Patterns
Cooperation and competition coexist:
- Market mechanisms: Agents achieve individual and collective benefits through trade
- Team competition: Intra-group cooperation, inter-group competition
Multi-Agent System Classification
graph TB
MAS[Multi-Agent Systems] --> STRUCT[By Structure]
MAS --> COMM[By Communication]
MAS --> GOAL[By Goal]
STRUCT --> FLAT[Flat<br/>Peer Agents]
STRUCT --> HIER[Hierarchical<br/>Manager-Worker]
STRUCT --> HYBRID[Hybrid]
COMM --> DIRECT[Direct Communication<br/>Message Passing]
COMM --> INDIRECT[Indirect Communication<br/>Blackboard/Environment]
COMM --> BROADCAST[Broadcast]
GOAL --> COOP[Cooperative<br/>Shared Goals]
GOAL --> COMP[Competitive<br/>Conflicting Goals]
GOAL --> MIXED[Mixed]
Why Multi-Agent Systems Are Needed
Limitations of Single Agents
- Capability bottleneck: A single agent cannot master all domains
- Context limitations: Complex tasks require context beyond a single agent's processing capacity
- Quality issues: Lack of external checks and feedback
- Robustness: Single point of failure risk
Advantages of Multiple Agents
- Specialization: Each agent focuses on its area of expertise
- Quality assurance: Agents cross-check and correct each other
- Parallel processing: Multiple agents work simultaneously to improve efficiency
- Emergent capabilities: Groups may exhibit abilities that individuals lack
Typical Application Scenarios
| Scenario | Agent Roles | Interaction Mode |
|---|---|---|
| Software development | Product manager, architect, developer, tester | Hierarchical cooperation |
| Research analysis | Researcher, analyst, writer, reviewer | Sequential cooperation |
| Debate-based reasoning | Proponent, opponent, judge | Adversarial competition |
| Creative generation | Brainstormer, evaluator, optimizer | Iterative cooperation |
| Customer service | Router, specialist, QA | Hierarchical cooperation |
Chapter Structure
- Communication Protocols and Message Passing - How agents communicate
- Collaboration Patterns - Common multi-agent collaboration architectures
- Game Theory and Mechanism Design - Mathematical foundations of competition and cooperation
- Emergent Behavior and Swarm Intelligence - Emergent properties of multi-agent systems
- Multi-Agent Frameworks - Practical frameworks and tools
- Multi-Agent Frontiers - Latest research directions
References
- Wooldridge, M., & Jennings, N. R. (1995). "Intelligent agents: Theory and practice"
- Guo, T., et al. (2024). "Large Language Model based Multi-Agents: A Survey of Progress and Challenges"
- Talebirad, Y., & Nadiri, A. (2023). "Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents"
- Li, G., et al. (2024). "A Survey on LLM-based Multi-Agent Systems"