Multi-Agent Patterns

Relevant source files

• 00-course-setup/README.md
• 01-intro-to-ai-agents/README.md
• 02-explore-agentic-frameworks/README.md
• 03-agentic-design-patterns/README.md
• 04-tool-use/README.md
• 05-agentic-rag/README.md
• 05-agentic-rag/code_samples/05-dotnet-agent-framework.cs
• 05-agentic-rag/code_samples/05-dotnet-agent-framework.ipynb
• 05-agentic-rag/code_samples/05-dotnet-agent-framework.md
• 06-building-trustworthy-agents/README.md
• 07-planning-design/README.md
• 08-multi-agent/README.md
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/01.dotnet-agent-framework-workflow-ghmodel-basic.cs
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/01.dotnet-agent-framework-workflow-ghmodel-basic.md
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/02.dotnet-agent-framework-workflow-ghmodel-sequential.cs
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/02.dotnet-agent-framework-workflow-ghmodel-sequential.md
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.md
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.cs
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.ipynb
• 08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.md
• 08-multi-agent/code_samples/workflows-agent-framework/python/01.python-agent-framework-workflow-ghmodel-basic.ipynb
• 08-multi-agent/code_samples/workflows-agent-framework/python/04.python-agent-framework-workflow-aifoundry-condition.ipynb
• 08-multi-agent/solution/solution-quiz.md
• 08-multi-agent/solution/solution.md
• 09-metacognition/README.md
• 10-ai-agents-production/README.md
• 14-microsoft-agent-framework/code-samples/14-concurrent.ipynb
• 14-microsoft-agent-framework/code-samples/14-conditional-workflow.ipynb
• 14-microsoft-agent-framework/code-samples/14-handoff.ipynb
• 14-microsoft-agent-framework/code-samples/14-human-loop.ipynb
• 14-microsoft-agent-framework/code-samples/14-middleware.ipynb
• 14-microsoft-agent-framework/code-samples/14-sequential.ipynb
• 14-microsoft-agent-framework/code-samples/hotel_booking_workflow_sample.py
• README.md

Purpose and Scope

This document details three fundamental multi-agent orchestration patterns: Group Chat (democratic collaboration), Hand-off (sequential processing), and Collaborative Filtering (parallel expertise). It explains the implementation, data flow, and key code entities used in the Microsoft Agent Framework (MAF) and AutoGen to facilitate complex collaboration between specialized agents.

For individual agent capabilities, see Tool Use Pattern (#4.1). For task decomposition, see Planning Pattern (#4.3).

Sources: 08-multi-agent/README.md1-28 01-intro-to-ai-agents/README.md102-116

When to Use Multi-Agent Systems

Multi-agent systems (MAS) are superior to monolithic single-agent designs when tasks involve:

• Large Workloads: Dividing tasks into smaller units for parallel processing 08-multi-agent/README.md36-37
• Complex Tasks: Breaking down intricate goals into subtasks assigned to specialized agents 08-multi-agent/README.md37-38
• Diverse Expertise: Leveraging specialized knowledge (e.g., separate agents for diagnostics, treatment, and monitoring in healthcare) 08-multi-agent/README.md38-39

Advantages

• Specialization: Prevents agent confusion by narrowing the scope of tools and instructions per agent 08-multi-agent/README.md43-45
• Scalability: Easier to add new agents than to overload a single agent's context and toolset 08-multi-agent/README.md45-46
• Fault Tolerance: If one agent fails, others can continue functioning 08-multi-agent/README.md46-47

Sources: 08-multi-agent/README.md32-50

Core Multi-Agent Orchestration Patterns

Group Chat Pattern (Democratic Collaboration)

In a Group Chat, multiple agents communicate in a shared conversation space. A manager or a predefined protocol (like Round Robin) determines the next speaker.

• Implementation: AutoGen uses RoundRobinGroupChat or GroupChatManager 02-explore-agentic-frameworks/README.md231-262
• Data Flow: Messages are broadcast to a shared history, allowing all agents to maintain state.

Sources: 08-multi-agent/README.md83-91 02-explore-agentic-frameworks/README.md231-262

Hand-off Pattern (Sequential Processing)

The Hand-off pattern transfers control from one agent to another once specific criteria are met. This is ideal for linear workflows like customer support escalation.

• Implementation: In MAF, this is achieved using WorkflowBuilder.AddEdge() to chain agents 08-multi-agent/code_samples/workflows-agent-framework/dotNET/02.dotnet-agent-framework-workflow-ghmodel-sequential.cs86-90
• Example: A "cancel trip" request split into sub-agents for flight, hotel, and car rental 01-intro-to-ai-agents/README.md71

Sources: 08-multi-agent/README.md93-101 14-microsoft-agent-framework/code-samples/14-handoff.ipynb

Collaborative Filtering (Parallel Expertise)

This pattern involves multiple agents working concurrently on the same problem from different perspectives, then synthesizing results.

• Implementation: MAF uses AddFanOutEdge() to broadcast and AddFanInEdge() to aggregate 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs58-61

Sources: 08-multi-agent/README.md103-115 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs58-61

Microsoft Agent Framework (MAF) Implementation Detail

MAF provides a structured WorkflowBuilder to define agent relationships.

Sequential Workflow

Implemented by chaining AddEdge calls. For example, a furniture sales workflow:

1. SalesAnalysisAgent
2. PriceEstimationAgent
3. QuoteGenerationAgent

08-multi-agent/code_samples/workflows-agent-framework/dotNET/02.dotnet-agent-framework-workflow-ghmodel-sequential.cs86-90

Concurrent Workflow (Fan-Out/Fan-In)

Requires two specialized executors:

• ConcurrentStartExecutor: Uses context.SendMessageAsync() to trigger multiple agents simultaneously 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs80-95
• ConcurrentAggregationExecutor: Collects responses into a list and yields a single aggregated ChatMessage 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs110-125

Conditional Workflow

Uses predicates to route data dynamically.

08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.cs160-165

Sources: 08-multi-agent/code_samples/workflows-agent-framework/dotNET/02.dotnet-agent-framework-workflow-ghmodel-sequential.cs1-126 08-multi-agent/code_samples/workflows-agent-framework/dotNET/03.dotnet-agent-framework-workflow-ghmodel-concurrent.cs58-127 08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.cs1-291

Composition Example: Refund Process

The refund process demonstrates pattern composition where specialized "Expertise Units" interact.

Expertise Units

Agent	Role	Pattern Usage
CustomerAgent	Initiator	Sequential (Start)
SellerAgent	Evaluator	Conditional (Approve/Reject)
ShippingAgent	Logistics	Parallel (with Payment)
PaymentAgent	Financials	Parallel (with Shipping)
ComplianceAgent	Oversight	Background Monitor

Data Flow Diagram

This diagram bridges Natural Language concepts to the WorkflowBuilder logic used in the codebase.

Sources: 08-multi-agent/README.md117-145 08-multi-agent/code_samples/workflows-agent-framework/dotNET/04.dotnet-agent-framework-workflow-aifoundry-condition.cs150-180

Visibility and Trust

Multi-Agent Visibility

Tracking interactions is essential for debugging. Observability tools (e.g., Langfuse, Microsoft Foundry) represent agent runs as Traces (complete tasks) and Spans (individual agent steps) 10-ai-agents-production/README.md22-26

Human-in-the-Loop

In complex multi-agent systems, agents must be instructed when to ask for human intervention (e.g., confirming a high-value refund) 08-multi-agent/README.md63-64 In AutoGen, this is managed by the UserProxyAgent 02-explore-agentic-frameworks/README.md280-285

Sources: 08-multi-agent/README.md65-78 10-ai-agents-production/README.md20-40 06-building-trustworthy-agents/README.md156-182