Multi-Agent Collaboration
Mechanisms: A Survey
A systematic taxonomy of interaction structures, communication topologies, consensus algorithms, and emergent pathologies in LLM multi-agent societies.
The Four Dimensions of Agent Collaboration
Every multi-agent LLM system can be decomposed along four fundamental design axes.
Information Routing
How messages travel between nodes: Centralized Star, Hierarchical Tree, Peer-to-Peer Mesh, or Linear Assembly Line.
Agent Incentives
Cooperative (shared reward), Competitive (zero-sum gaming), or Adversarial Debate (truth discovery through thesis-antithesis).
Decision Synthesis
How the collective picks a final answer: Majority Voting, Elo Tournament ranking, Hierarchical Judge, or Iterative Discussion.
State Sharing
Private local scratchpads vs Central Blackboards vs Pub-Sub message buses with topic filters.
Visualizing Communication Networks
Select a network topology to compare message complexity, fault tolerance, and context token overhead.
Multi-Agent Emergent Failure Modes
When multi-agent societies malfunction, they exhibit failure patterns not seen in single-agent prompts.
Echo Chambers
Agents quickly converge to agree with the most confident or first-responding agent, abandoning correct contrarian factual knowledge.
Chinese Whispers
Across long sequential chains, critical constraints or numerical bounds get dropped or subtly distorted as summaries replace raw evidence.
Politeness Deadlocks
Without strict halting criteria, agents enter cycles of mutual praise and repetitive meta-commentary, draining API budgets.
Collaboration Mechanisms Quiz
Test your understanding of multi-agent topologies and collective coordination strategies.