Multi-Agent Systems

Multi-Agent Systems (MAS) distribute tasks among specialized agents with distinct roles, improving modularity, scalability, and performance.

Core Architectures

• Orchestrator/Manager Pattern: A primary agent (the Manager) delegates sub-tasks to specialized worker agents (e.g., Search Agent, Code Interpreter Agent).
• Handoff Pattern: Agents transfer control to another agent better suited for the current task (common in LlamaIndex AgentWorkflow).
• Hierarchical Team: A tree structure where managers coordinate teams of workers.

Benefits

1. Focus: Each agent has a narrower scope and fewer tools, reducing token bloat and latency.
2. Robustness: Errors in one sub-task can be caught and corrected by the manager without failing the entire request.
3. Scalability: Complex tasks (e.g., "Plan a party and generate a map") are broken into manageable chunks.

Implementation Examples

• smolagents: Uses ManagedAgent to wrap workers.
• LlamaIndex: Uses AgentWorkflow to handle multi-agent loops and handoffs.
• openai-symphony: A service spec where the orchestrator manages coding-agent workspaces and issue lifecycle.

See Also

• agentic-frameworks-moc
• smolagents
• llamaindex