Multi Agent Orchestration

Multi-Agent Orchestration

Design and orchestrate sophisticated multi-agent systems where specialized agents collaborate to solve complex problems, combining different expertise and perspectives.

Quick Start

Get started with multi-agent implementations in the examples and utilities:

• Examples: See [examples/](examples/) directory for complete implementations:
• [orchestration_patterns.py](examples/orchestration_patterns.py) - Sequential, parallel, hierarchical, and consensus orchestration
• [framework_implementations.py](examples/framework_implementations.py) - Templates for CrewAI, AutoGen, LangGraph, and Swarm

• Utilities: See [scripts/](scripts/) directory for helper modules:
• [agent_communication.py](scripts/agent_communication.py) - Message broker, shared memory, and communication protocols
• [workflow_management.py](scripts/workflow_management.py) - Workflow execution, optimization, and monitoring
• [benchmarking.py](scripts/benchmarking.py) - Team performance and agent effectiveness metrics

Overview

Multi-agent systems decompose complex problems into specialized sub-tasks, assigning each to an agent with relevant expertise, then coordinating their work toward a unified goal.

When Multi-Agent Systems Shine

• Complex Workflows: Tasks requiring multiple specialized roles
• Domain-Specific Expertise: Finance, legal, HR, engineering need different knowledge
• Parallel Processing: Multiple agents work on different aspects simultaneously
• Collaborative Reasoning: Agents debate, refine, and improve solutions
• Resilience: Failures in one agent don't break the entire system
• Scalability: Easy to add new specialized agents

Architecture Overview

User Request
    ↓
Orchestrator
    ├→ Agent 1 (Specialist) → Task 1
    ├→ Agent 2 (Specialist) → Task 2
    ├→ Agent 3 (Specialist) → Task 3
    ↓
Result Aggregator
    ↓
Final Response

Core Concepts

Agent Definition

An agent is defined by:

• Role: What responsibility does it have? (e.g., "Financial Analyst")
• Goal: What should it accomplish? (e.g., "Analyze financial risks")
• Expertise: What knowledge/tools does it have?
• Tools: What capabilities can it access?
• Context: What information does it need to work effectively?

Orchestration Patterns

1. Sequential Orchestration

• Agents work one after another
• Each agent uses output from previous agent
• Use Case: Steps must follow order (research → analysis → writing)

2. Parallel Orchestration

• Multiple agents work simultaneously
• Results aggregated at the end
• Use Case: Independent tasks (analyze competitors, market, users)

3. Hierarchical Orchestration

• Senior agent delegates to junior agents
• Manager coordinates flow
• Use Case: Large projects with oversight

4. Consensus-Based Orchestration

• Multiple agents analyze problem
• Debate and refine ideas
• Vote or reach consensus
• Use Case: Complex decisions needing multiple perspectives

5. Tool-Mediated Orchestration

• Agents use shared tools/databases
• Minimal direct communication
• Use Case: Large systems, indirect coordination

Multi-Agent Team Examples

Finance Team

Coordinator Agent
    ├→ Market Analyst Agent
    │   ├ Tools: Market data API, financial news
    │   └ Task: Analyze market conditions
    ├→ Financial Analyst Agent
    │   ├ Tools: Financial statements, ratio calculations
    │   └ Task: Analyze company financials
    ├→ Risk Manager Agent
    │   ├ Tools: Risk models, scenario analysis
    │   └ Task: Assess investment risks
    └→ Report Writer Agent
        ├ Tools: Document generation
        └ Task: Synthesize findings into report

Legal Team

Case Manager Agent (Coordinator)
    ├→ Contract Analyzer Agent
    │   └ Task: Review contract terms
    ├→ Precedent Research Agent
    │   └ Task: Find relevant case law
    ├→ Risk Assessor Agent
    │   └ Task: Identify legal risks
    └→ Document Drafter Agent
        └ Task: Prepare legal documents

Customer Support Team

Support Coordinator
    ├→ Issue Classifier Agent
    │   └ Task: Categorize customer issue
    ├→ Knowledge Base Agent
    │   └ Task: Find relevant documentation
    ├→ Escalation Agent
    │   └ Task: Determine if human escalation needed
    └→ Solution Synthesizer Agent
        └ Task: Prepare comprehensive response

Implementation Frameworks

1. CrewAI

Best For: Teams with clear roles and hierarchical structure

from crewai import Agent, Task, Crew

# Define agents
analyst = Agent(
    role="Financial Analyst",
    goal="Analyze financial data and provide insights",
    backstory="Expert in financial markets with 10+ years experience"
)

researcher = Agent(
    role="Market Researcher",
    goal="Research market trends and competition",
    backstory="Data-driven researcher specializing in market analysis"
)

# Define tasks
analysis_task = Task(
    description="Analyze Q3 financial results for {company}",
    agent=analyst,
    tools=[financial_tool, data_tool]
)

research_task = Task(
    description="Research competitive landscape in {market}",
    agent=researcher,
    tools=[web_search_tool, industry_data_tool]
)

# Create crew and execute
crew = Crew(
    agents=[analyst, researcher],
    tasks=[analysis_task, research_task],
    process=Process.sequential
)

result = crew.kickoff(inputs={"company": "TechCorp", "market": "AI"})

2. AutoGen (Microsoft)

Best For: Complex multi-turn conversations and negotiations

from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager

# Define agents
analyst = AssistantAgent(
    name="analyst",
    system_message="You are a financial analyst..."
)

researcher = AssistantAgent(
    name="researcher",
    system_message="You are a market researcher..."
)

# Create group chat
groupchat = GroupChat(
    agents=[analyst, researcher],
    messages=[],
    max_round=10,
    speaker_selection_method="auto"
)

# Manage group conversation
manager = GroupChatManager(groupchat=groupchat)

# User proxy to initiate conversation
user = UserProxyAgent(name="user")

# Have conversation
user.initiate_chat(
    manager,
    message="Analyze if Company X should invest in Y market"
)

3. LangGraph

Best For: Complex workflows with state management

from langgraph.graph import Graph, StateGraph
from langgraph.prebuilt import create_agent_executor

# Define state
class AgentState:
    research_findings: str
    analysis: str
    recommendations: str

# Create graph
graph = StateGraph(AgentState)

# Add nodes for each agent
graph.add_node("researcher", research_agent)
graph.add_node("analyst", analyst_agent)
graph.add_node("writer", writer_agent)

# Define edges (workflow)
graph.add_edge("researcher", "analyst")
graph.add_edge("analyst", "writer")

# Set entry/exit points
graph.set_entry_point("researcher")
graph.set_finish_point("writer")

# Compile and run
workflow = graph.compile()
result = workflow.invoke({"topic": "AI trends"})

4. OpenAI Swarm

Best For: Simple agent handoffs and conversational workflows

from swarm import Agent, Swarm

# Define agents
triage_agent = Agent(
    name="Triage Agent",
    instructions="Determine which specialist to route the customer to"
)

billing_agent = Agent(
    name="Billing Specialist",
    instructions="Handle billing and payment questions"
)

technical_agent = Agent(
    name="Technical Support",
    instructions="Handle technical issues"
)

# Define handoff functions
def route_to_billing(reason: str):
    return billing_agent

def route_to_technical(reason: str):
    return technical_agent

# Add tools to triage agent
triage_agent.functions = [route_to_billing, route_to_technical]

# Execute swarm
client = Swarm()
response = client.run(
    agent=triage_agent,
    messages=[{"role": "user", "content": "I have a billing question"}]
)

Orchestration Patterns

Pattern 1: Sequential Task Chain

Agents execute tasks in sequence, each building on previous results:

# Task 1: Research
research_output = research_agent.work("Analyze AI market trends")

# Task 2: Analysis (uses research output)
analysis = analyst_agent.work(f"Analyze these findings: {research_output}")

# Task 3: Report (uses analysis)
report = writer_agent.work(f"Write report on: {analysis}")

When to Use: Steps have dependencies, each builds on previous

Pattern 2: Parallel Execution

Multiple agents work simultaneously, results combined:

import asyncio

async def parallel_teams():
    # All agents work in parallel
    market_task = market_agent.work_async("Analyze market")
    technical_task = tech_agent.work_async("Analyze technology")
    user_task = user_agent.work_async("Analyze user needs")

    # Wait for all to complete
    market_results, tech_results, user_results = await asyncio.gather(
        market_task, technical_task, user_task
    )

    # Synthesize results
    return synthesize(market_results, tech_results, user_results)

When to Use: Independent analyses, need quick results, want diversity

Pattern 3: Hierarchical Structure

Manager agent coordinates specialists:

manager_agent.orchestrate({
    "market_analysis": {
        "agents": [competitor_analyst, trend_analyst],
        "task": "Comprehensive market analysis"
    },
    "technical_evaluation": {
        "agents": [architecture_agent, security_agent],
        "task": "Technical feasibility assessment"
    },
    "synthesis": {
        "agents": [strategy_agent],
        "task": "Create strategic recommendations"
    }
})

When to Use: Clear hierarchy, different teams, complex coordination

Pattern 4: Debate & Consensus

Multiple agents discuss and reach consensus:

agents = [bull_agent, bear_agent, neutral_agent]
question = "Should we invest in this startup?"

# Debate round 1
arguments = {agent: agent.argue(question) for agent in agents}

# Debate round 2 (respond to others)
counter_arguments = {
    agent: agent.respond(arguments) for agent in agents
}

# Reach consensus
consensus = mediator_agent.synthesize_consensus(counter_arguments)

When to Use: Complex decisions, need multiple perspectives, risk assessment

Agent Communication Patterns

1. Direct Communication

Agents pass messages directly to each other:

agent_a.send_message(agent_b, {
    "type": "request",
    "action": "analyze_document",
    "document": doc_content,
    "context": {"deadline": "urgent"}
})

2. Tool-Mediated Communication

Agents use shared tools/databases:

# Agent A writes to shared memory
shared_memory.write("findings", {"market_size": "$5B", "growth": "20%"})

# Agent B reads from shared memory
findings = shared_memory.read("findings")

3. Manager-Based Communication

Central coordinator manages agent communication:

manager.broadcast("update_all_agents", {
    "new_deadline": "tomorrow",
    "priority": "critical"
})

Best Practices

Agent Design

• ✓ Clear, specific role and goal
• ✓ Appropriate tools for the role
• ✓ Relevant background/expertise
• ✓ Distinct from other agents
• ✓ Reasonable scope of work

Workflow Design

• ✓ Clear task dependencies
• ✓ Identified handoff points
• ✓ Error handling between agents
• ✓ Fallback strategies
• ✓ Performance monitoring

Communication

• ✓ Structured message formats
• ✓ Clear context sharing
• ✓ Error propagation strategy
• ✓ Timeout handling
• ✓ Audit logging

Orchestration

• ✓ Define process clearly (sequential, parallel, etc.)
• ✓ Set clear success criteria
• ✓ Monitor agent performance
• ✓ Implement feedback loops
• ✓ Allow human intervention points

Common Challenges & Solutions

Challenge: Agent Conflicts

Solutions:

• Clear role separation
• Explicit decision-making rules
• Consensus mechanisms
• Conflict resolution agent
• Clear authority hierarchy

Challenge: Slow Execution

Solutions:

• Use parallel execution where possible
• Cache results from expensive operations
• Pre-process data
• Optimize agent logic
• Implement timeout handling

Challenge: Poor Quality Results

Solutions:

• Better agent prompts/instructions
• More relevant tools
• Feedback integration
• Quality validation agents
• Result aggregation strategies

Challenge: Complex Workflows

Solutions:

• Break into smaller teams
• Hierarchical structure
• Clear task definitions
• Good state management
• Documentation of workflow

Evaluation Metrics

Team Performance:

• Task completion rate
• Quality of results
• Execution time
• Cost (tokens/API calls)
• Error rate

Agent Effectiveness:

• Task success rate
• Response quality
• Tool usage efficiency
• Communication clarity
• Collaboration score

Advanced Techniques

1. Self-Organizing Teams

Agents autonomously decide roles and workflow:

# Agents negotiate roles based on task
agents = [agent1, agent2, agent3]
task = "complex financial analysis"

# Agents determine best structure
negotiated_structure = self_organize(agents, task)
# Returns optimal workflow for this task

2. Adaptive Workflows

Workflow changes based on progress:

# Monitor progress
if progress < expected_rate:
    # Increase resources
    workflow.add_agent(specialist_agent)
elif quality < threshold:
    # Increase validation
    workflow.insert_review_step()

3. Cross-Agent Learning

Agents learn from each other's work:

# After team execution
execution_trace = crew.get_execution_trace()

# Extract learnings
learnings = extract_patterns(execution_trace)

# Update agent knowledge
for agent, learning in learnings.items():
    agent.update_knowledge(learning)

Resources

Frameworks

• CrewAI: https://crewai.com/
• AutoGen: https://microsoft.github.io/autogen/
• LangGraph: https://langchain-ai.github.io/langgraph/
• Swarm: https://github.com/openai/swarm

Papers

• "Generative Agents" (Park et al.)
• "Self-Organizing Multi-Agent Systems" (research papers)

Implementation Checklist

• [ ] Define each agent's role, goal, and expertise
• [ ] Identify available tools/capabilities for each agent
• [ ] Plan workflow (sequential, parallel, hierarchical)
• [ ] Define communication patterns
• [ ] Implement task definitions
• [ ] Set success criteria for each task
• [ ] Add error handling and fallbacks
• [ ] Implement monitoring/logging
• [ ] Test team collaboration
• [ ] Evaluate quality and performance
• [ ] Optimize based on results
• [ ] Document workflow and decisions

Getting Started

1. Start Small: Begin with 2-3 agents
2. Clear Workflow: Document how agents interact
3. Test Thoroughly: Validate agent behavior individually and together
4. Monitor Closely: Track performance and results
5. Iterate: Refine based on results
6. Scale: Add agents and complexity as needed

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: javalenciacai
• Source: javalenciacai/QASkills
• License: MIT

Install and usage instructions live in the source repository linked above.