multi-agent-architect（多 Agent 架构师）

Multi-Agent Architect

Design systems where multiple specialized agents collaborate to solve complex problems.

Core Principle

Divide complex tasks among specialized agents, each expert in their domain, coordinated through clear communication patterns.

When to Use Multi-Agent Systems

Use Multi-Agent When:

• ✅ Task requires multiple specializations (research + writing + coding)
• ✅ Parallel processing speeds up solution (independent subtasks)
• ✅ Need self-correction through peer review
• ✅ Complex workflows with decision points
• ✅ Scaling single-agent becomes unwieldy

Don't Use Multi-Agent When:

• ❌ Single agent can handle task efficiently
• ❌ Task is simple and linear
• ❌ Communication overhead > parallelization benefit
• ❌ Team lacks multi-agent debugging expertise

Multi-Agent Patterns

Pattern 1: Sequential Pipeline

Use: Multi-step workflow where each agent builds on previous

User Query → Researcher → Analyst → Writer → Editor → Output

Example: Research report generation

1. Researcher: Gather sources
2. Analyst: Synthesize findings
3. Writer: Draft report
4. Editor: Refine and format

Pros: Clear dependencies, easy to debug Cons: Sequential (no parallelization), bottlenecks

Pattern 2: Hierarchical (Manager-Worker)

Use: Complex task broken into parallel subtasks

              Manager Agent
              /     |     \
    Worker 1   Worker 2   Worker 3
    (Search)   (Analyze)  (Summarize)
              \     |     /
              Aggregator Agent

Example: Market research across competitors

• Manager: Decompose into per-competitor analysis
• Workers: Research competitor A, B, C in parallel
• Aggregator: Combine findings

Pros: Parallelization, specialization Cons: Manager complexity, coordination overhead

Pattern 3: Peer Collaboration (Round Table)

Use: Multiple perspectives improve quality

Coder ↔ Reviewer ↔ Tester
  ↓        ↓        ↓
      Consensus

Example: Code generation with review

1. Coder: Write initial code
2. Reviewer: Check for issues
3. Tester: Validate functionality
4. Iterate until consensus

Pros: Quality through review, self-correction Cons: May not converge, expensive (multiple LLM calls)

Pattern 4: Agent Swarm

Use: Many agents explore solution space independently

Agent 1 → Candidate Solution 1
Agent 2 → Candidate Solution 2
Agent 3 → Candidate Solution 3
   ↓
Selector (pick best)

Example: Creative brainstorming

• 5 agents generate different approaches
• Selector evaluates and picks best

Pros: Exploration, creativity Cons: Cost (N agents), may produce similar solutions

Communication Patterns

1. Shared Memory

shared_state = {
    "research_findings": [],
    "current_task": "analyze_competitors",
    "decisions": []
}

# All agents read/write to shared state
researcher.execute(shared_state)
analyst.execute(shared_state)

Pros: Simple, all agents see full context Cons: Race conditions, hard to debug who changed what

2. Message Passing

# Agent A sends message to Agent B
message = {
    "from": "researcher",
    "to": "analyst",
    "content": research_findings,
    "metadata": {"confidence": 0.9}
}

message_queue.send(message)

Pros: Clear communication flow, traceable Cons: More complex to implement

3. Event-Driven

# Agents subscribe to events
event_bus.subscribe("research_complete", analyst.on_research_complete)
event_bus.subscribe("analysis_complete", writer.on_analysis_complete)

# Agent publishes event when done
event_bus.publish("research_complete", research_data)

Pros: Loose coupling, scalable Cons: Harder to follow execution flow

Agent Coordination Strategies

1. Fixed Workflow

Predefined sequence, no dynamic decisions

workflow = [
    ("researcher", gather_info),
    ("analyst", analyze_data),
    ("writer", create_report)
]

for agent_name, task in workflow:
    result = agents[agent_name].execute(task, context)
    context.update(result)

Use: Predictable tasks, clear dependencies

2. Dynamic Routing

Manager decides next agent based on context

class ManagerAgent:
    def route_task(self, task, context):
        if requires_technical_expertise(task):
            return tech_specialist
        elif requires_creative_input(task):
            return creative_agent
        else:
            return generalist

Use: Tasks vary significantly, need flexibility

3. Consensus-Based

Agents vote or reach agreement

proposals = [agent.propose_solution(task) for agent in agents]
scores = [agent.evaluate(proposals) for agent in agents]
best = proposals[argmax(mean(scores))]

Use: High-stakes decisions, quality critical

Implementation with CrewAI

CrewAI Pattern (Role-based teams):

from crewai import Agent, Task, Crew

# Define specialized agents
researcher = Agent(
    role="Research Specialist",
    goal="Gather comprehensive information on {topic}",
    backstory="Expert researcher with 10 years experience",
    tools=[search_tool, scrape_tool]
)

analyst = Agent(
    role="Data Analyst",
    goal="Synthesize research findings into insights",
    backstory="Data scientist specialized in trend analysis",
    tools=[analysis_tool]
)

writer = Agent(
    role="Technical Writer",
    goal="Create clear, compelling reports",
    backstory="Professional writer with technical expertise",
    tools=[writing_tool]
)

# Define tasks
research_task = Task(
    description="Research {topic} thoroughly",
    agent=researcher,
    expected_output="Comprehensive research findings with sources"
)

analysis_task = Task(
    description="Analyze research findings for key insights",
    agent=analyst,
    context=[research_task],  # Depends on research_task
    expected_output="List of key insights and trends"
)

writing_task = Task(
    description="Write executive summary based on analysis",
    agent=writer,
    context=[research_task, analysis_task],
    expected_output="500-word executive summary"
)

# Create crew and execute
crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[research_task, analysis_task, writing_task],
    verbose=True
)

result = crew.kickoff(inputs={"topic": "AI market trends"})

Implementation with LangGraph

LangGraph Pattern (State machines):

from langgraph.graph import StateGraph, END

class AgentState(TypedDict):
    input: str
    research: str
    analysis: str
    output: str

def research_node(state):
    research = researcher_agent.run(state["input"])
    return {"research": research}

def analysis_node(state):
    analysis = analyst_agent.run(state["research"])
    return {"analysis": analysis}

def writing_node(state):
    output = writer_agent.run(state["analysis"])
    return {"output": output}

# Build graph
workflow = StateGraph(AgentState)

workflow.add_node("research", research_node)
workflow.add_node("analysis", analysis_node)
workflow.add_node("writing", writing_node)

workflow.set_entry_point("research")
workflow.add_edge("research", "analysis")
workflow.add_edge("analysis", "writing")
workflow.add_edge("writing", END)

app = workflow.compile()

# Execute
result = app.invoke({"input": "Analyze AI market trends"})

Best Practices

1. Clear Agent Roles

Each agent should have specific expertise and responsibilities

2. Minimize Communication

More agents = more coordination overhead. Start simple.

3. Idempotent Operations

Agents should be restartable without side effects

4. Failure Handling

Design for agent failures (retry, fallback, skip)

5. Observable Execution

Log agent decisions, trace execution flow

6. Cost Management

Track token usage per agent, optimize expensive calls

Common Multi-Agent Mistakes

❌ Too many agents → Start with 2-3, add only if needed ❌ Unclear responsibilities → Define explicit roles ❌ No failure handling → One agent failure breaks entire system ❌ Synchronous bottlenecks → Parallelize independent agents ❌ Ignoring costs → N agents = N× LLM calls ❌ Over-engineering → Single agent often sufficient

Decision Framework: Single vs Multi-Agent

Task Complexity?
│
├─ Simple, linear → Single Agent
│
├─ Complex, requires specialization?
│  │
│  ├─ Sequential steps → Pipeline Pattern
│  ├─ Parallel subtasks → Hierarchical Pattern
│  ├─ Need review → Peer Collaboration
│  └─ Explore solutions → Swarm Pattern
│
└─ Uncertain → Start with Single Agent, refactor to Multi if needed

Monitoring & Debugging

# Track agent execution
class TrackedAgent(Agent):
    def execute(self, task, context):
        start = time.time()
        logger.info(f"{self.name} starting: {task}")

        result = super().execute(task, context)

        duration = time.time() - start
        logger.info(f"{self.name} completed in {duration}s")

        metrics.record({
            "agent": self.name,
            "task": task,
            "duration": duration,
            "tokens": result.token_count,
            "cost": result.cost
        })

        return result

Key Metrics:

• Agent execution time
• Token usage per agent
• Success/failure rates
• Handoff delays
• Overall workflow duration

Related Resources

Related Skills:

• rag-implementer - For knowledge-grounded agents
• knowledge-graph-builder - For agent knowledge bases
• api-designer - For agent communication APIs

Related Patterns:

• META/DECISION-FRAMEWORK.md - Framework selection (CrewAI vs LangGraph)
• STANDARDS/architecture-patterns/multi-agent-pattern.md - Agent architectures (when created)

Related Playbooks:

• PLAYBOOKS/deploy-multi-agent-system.md - Deployment guide (when created)
• PLAYBOOKS/debug-agent-workflows.md - Debugging procedures (when created)