Agent Orchestration

Agent Orchestration: Multi-Agent System Patterns

Core Principles

Complex problems often require multiple specialized agents working together. ADK provides robust patterns for orchestrating agent teams, enabling modular, maintainable, and scalable agentic systems.

Key Insight: Single monolithic agents with overly complex prompts are harder to maintain and debug than teams of focused specialists coordinated by a simple orchestrator.

Agent Types: LlmAgent vs WorkflowAgent

LlmAgent (Dynamic Reasoning)

Purpose: For tasks where the next action depends on runtime reasoning and context.

Characteristics:

• LLM decides which tool to call and when
• Flexible, adaptive behavior
• Suitable for open-ended problems
• Can handle unexpected inputs

Use Cases:

• Conversational assistants
• Complex problem-solving requiring judgment
• Tasks with unpredictable user requests
• Research and analysis workflows

Example:

from google import genai
from google.genai import types

async def create_research_agent() -> types.Agent:
    """
    Create LlmAgent for research tasks.

    The agent dynamically decides whether to search, read, or summarize
    based on user questions and intermediate findings.
    """
    client = genai.Client(vertexai=True)

    # Define tools
    search_tool = types.Tool(function_declarations=[...])
    read_tool = types.Tool(function_declarations=[...])
    summarize_tool = types.Tool(function_declarations=[...])

    # LlmAgent with dynamic tool selection
    agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        You are a research assistant. For each user query:
        1. Use search_tool to find relevant sources
        2. Use read_tool to examine source content
        3. Use summarize_tool to synthesize findings

        Adapt your approach based on the query complexity.
        """,
        tools=[search_tool, read_tool, summarize_tool]
    )

    return agent

WorkflowAgent (Deterministic Flow)

Purpose: For tasks with predictable, repeatable processes where the execution flow is known in advance.

Characteristics:

• Hardcoded execution sequence
• Predictable, reliable behavior
• No LLM reasoning overhead for flow control
• Ideal for automation pipelines

Types:

• SequentialAgent: Execute agents one after another
• ParallelAgent: Execute agents concurrently
• LoopAgent: Repeat agent execution with conditions

Use Cases:

• Data processing pipelines
• Validation workflows
• Multi-stage transformations
• Scheduled automation tasks

Example - SequentialAgent:

from google.genai import types

async def create_document_processor() -> types.SequentialAgent:
    """
    Sequential workflow for document processing.

    Flow: Upload → Parse → Validate → Store (deterministic sequence)
    """
    # Define sub-agents for each stage
    upload_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Validate and upload documents to storage.",
        tools=[upload_tool]
    )

    parse_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Extract structured data from documents.",
        tools=[parse_tool]
    )

    validate_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Validate extracted data against schema.",
        tools=[validate_tool]
    )

    store_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Store validated data in database.",
        tools=[store_tool]
    )

    # Sequential workflow
    workflow = types.SequentialAgent(
        agents=[upload_agent, parse_agent, validate_agent, store_agent]
    )

    return workflow

Example - ParallelAgent:

async def create_parallel_analyzer() -> types.ParallelAgent:
    """
    Parallel analysis workflow.

    Run sentiment analysis, entity extraction, and summarization
    simultaneously for speed.
    """
    sentiment_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Analyze sentiment of the text.",
        tools=[sentiment_tool]
    )

    entity_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Extract named entities from text.",
        tools=[entity_tool]
    )

    summary_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="Generate concise summary of text.",
        tools=[summary_tool]
    )

    # Parallel execution (faster)
    parallel_workflow = types.ParallelAgent(
        agents=[sentiment_agent, entity_agent, summary_agent]
    )

    return parallel_workflow

Decision Matrix: Which Agent Type?

Coordinator Pattern (Recommended Architecture)

Root Coordinator with Specialist Sub-Agents

Architecture: Single coordinator agent dispatches tasks to specialized agents based on request type.

Benefits:

• Clear separation of concerns
• Easy to add new specialists
• Simple routing logic
• Improved debuggability

Implementation:

from google import genai
from google.genai import types
from pydantic import BaseModel, ConfigDict, Field

# Specialist agents
async def create_code_specialist() -> types.LlmAgent:
    """Agent specialized in code generation and review."""
    return types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        You are a senior software engineer specializing in Python.

        Your responsibilities:
        - Generate production-ready code with type hints
        - Review code for bugs and improvements
        - Suggest optimal algorithms and data structures

        Always follow PEP 8 and include comprehensive docstrings.
        """,
        tools=[code_analyzer_tool, code_formatter_tool]
    )

async def create_architecture_specialist() -> types.LlmAgent:
    """Agent specialized in system architecture."""
    return types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        You are a principal architect with 15+ years experience.

        Your responsibilities:
        - Design scalable system architectures
        - Evaluate trade-offs between approaches
        - Create architecture decision records (ADRs)

        Focus on maintainability, scalability, and security.
        """,
        tools=[diagram_tool, adr_tool]
    )

async def create_test_specialist() -> types.LlmAgent:
    """Agent specialized in testing."""
    return types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        You are a test automation engineer.

        Your responsibilities:
        - Generate comprehensive test suites
        - Design test strategies (unit, integration, e2e)
        - Achieve 80%+ code coverage

        Use pytest patterns and AAA structure.
        """,
        tools=[test_generator_tool, coverage_tool]
    )

# Coordinator agent
async def create_coordinator() -> types.LlmAgent:
    """
    Root coordinator that delegates to specialists.

    Analyzes user requests and routes to appropriate specialist.
    """
    # Create specialist agents
    code_agent = await create_code_specialist()
    arch_agent = await create_architecture_specialist()
    test_agent = await create_test_specialist()

    # Wrap specialists as tools (agent-as-tool pattern)
    code_tool = types.AgentTool(
        agent=code_agent,
        name="code_specialist",
        description="For code generation, review, and optimization tasks"
    )

    arch_tool = types.AgentTool(
        agent=arch_agent,
        name="architecture_specialist",
        description="For system design, architecture decisions, and scalability analysis"
    )

    test_tool = types.AgentTool(
        agent=test_agent,
        name="test_specialist",
        description="For test generation, test strategy, and coverage analysis"
    )

    # Coordinator with specialist tools
    coordinator = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        You are a project coordinator managing specialist engineers.

        Your role:
        1. Analyze user requests
        2. Determine which specialist(s) to consult
        3. Delegate tasks using appropriate specialist tools
        4. Synthesize results into cohesive response

        Available specialists:
        - code_specialist: For implementation and code review
        - architecture_specialist: For system design
        - test_specialist: For testing and quality assurance

        Routing rules:
        - "Create/implement/write code" → code_specialist
        - "Design/architect/plan system" → architecture_specialist
        - "Test/coverage/quality" → test_specialist
        - Complex tasks may require multiple specialists
        """,
        tools=[code_tool, arch_tool, test_tool]
    )

    return coordinator

Using the Coordinator

async def main() -> None:
    """Example usage of coordinator pattern."""
    coordinator = await create_coordinator()
    client = genai.Client(vertexai=True)

    # User request
    user_query = "Design and implement a user authentication system"

    # Coordinator analyzes and delegates
    # 1. Calls architecture_specialist to design system
    # 2. Calls code_specialist to implement
    # 3. Calls test_specialist to generate tests
    # 4. Synthesizes results into complete response

    response = await client.aio.chats.create(
        agent=coordinator
    ).send_message(user_query)

    print(response.text)

Inter-Agent Communication via Session State

Sharing Data Between Agents

Pattern: Use session state to pass context and results between agents in a workflow.

Example:

from google.genai import types

async def data_extraction_agent(context: types.ToolContext) -> dict:
    """
    Extract data from document and save to session state.

    Args:
        context: Tool execution context with session state

    Returns:
        Extracted data dictionary
    """
    # Extract data (implement your logic)
    extracted_data = {
        "entities": ["Alice", "Bob", "TechCorp"],
        "dates": ["2025-01-15"],
        "amounts": [150000]
    }

    # Save to session state for next agent
    context.state["extracted_data"] = extracted_data

    return extracted_data

async def data_validation_agent(context: types.ToolContext) -> dict:
    """
    Validate data extracted by previous agent.

    Args:
        context: Tool execution context with session state

    Returns:
        Validation results
    """
    # Retrieve data from session state
    extracted_data = context.state.get("extracted_data", {})

    # Validate data
    validation_results = {
        "valid": True,
        "errors": [],
        "warnings": ["Amount exceeds typical range"]
    }

    # Update session state
    context.state["validation_results"] = validation_results

    return validation_results

async def data_storage_agent(context: types.ToolContext) -> dict:
    """
    Store validated data in database.

    Args:
        context: Tool execution context with session state

    Returns:
        Storage confirmation
    """
    # Retrieve from session state
    extracted_data = context.state.get("extracted_data")
    validation_results = context.state.get("validation_results")

    # Only store if validation passed
    if validation_results.get("valid"):
        # Store data (implement your logic)
        return {"stored": True, "record_id": "12345"}
    else:
        return {"stored": False, "reason": "Validation failed"}

Best Practices:

• Use session state for lightweight data (< 10KB)
• Clear state after workflow completion
• Document state schema in agent instructions
• Validate state exists before accessing

Agent-as-Tool Pattern (AgentTool)

Wrapping Agents as Tools

Purpose: Make entire agents callable by other agents, enabling hierarchical delegation.

Pattern:

from google.genai import types

# Create specialist agent
specialist_agent = types.LlmAgent(
    model="gemini-2.0-flash-exp",
    system_instruction="You are a database optimization specialist...",
    tools=[query_analyzer_tool, index_advisor_tool]
)

# Wrap as tool
db_specialist_tool = types.AgentTool(
    agent=specialist_agent,
    name="database_optimizer",
    description="Optimize database queries and schema design. Use for slow queries, indexing strategy, or schema refactoring."
)

# Coordinator can now call the specialist
coordinator = types.LlmAgent(
    model="gemini-2.0-flash-exp",
    system_instruction="You coordinate technical specialists...",
    tools=[db_specialist_tool, other_tools...]
)

Benefits:

• Specialists maintain their own context and tools
• Coordinator stays simple (routing logic only)
• Easy to test specialists in isolation
• Specialists can be reused across coordinators

When to Use:

• Complex domain requiring specialized expertise
• Agent has multiple related tools
• Task requires extended reasoning in specific domain
• Want to isolate and test sub-functionality

Advanced Orchestration Patterns

Dynamic Agent Selection

async def create_dynamic_coordinator() -> types.LlmAgent:
    """
    Coordinator that dynamically selects agents based on query analysis.
    """
    # Create diverse specialists
    specialists = {
        "python": await create_python_specialist(),
        "typescript": await create_typescript_specialist(),
        "rust": await create_rust_specialist(),
        "devops": await create_devops_specialist(),
        "security": await create_security_specialist()
    }

    # Wrap all as tools
    specialist_tools = [
        types.AgentTool(
            agent=agent,
            name=name,
            description=f"{name.title()} specialist for {name}-related tasks"
        )
        for name, agent in specialists.items()
    ]

    coordinator = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        Analyze user queries and select appropriate specialist(s):

        Query indicators:
        - "Python"/"Django"/"Flask" → python specialist
        - "TypeScript"/"React"/"Node" → typescript specialist
        - "Rust"/"performance"/"systems" → rust specialist
        - "CI/CD"/"deployment"/"Docker" → devops specialist
        - "security"/"auth"/"encryption" → security specialist

        For cross-cutting concerns, consult multiple specialists.
        """,
        tools=specialist_tools
    )

    return coordinator

Multi-Stage Pipeline with Error Handling

async def create_robust_pipeline() -> types.SequentialAgent:
    """
    Pipeline with error handling and rollback capabilities.
    """
    # Stage 1: Input validation
    validation_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        Validate input data. Save validation results to session state.
        If validation fails, set state["pipeline_stop"] = True.
        """,
        tools=[validation_tool]
    )

    # Stage 2: Data transformation
    transform_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        Check state["pipeline_stop"]. If True, skip transformation.
        Otherwise, transform data and save to state.
        """,
        tools=[transform_tool]
    )

    # Stage 3: Storage
    storage_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        Check state["pipeline_stop"]. If True, skip storage.
        Otherwise, store data and record transaction ID.
        If storage fails, set state["rollback_needed"] = True.
        """,
        tools=[storage_tool]
    )

    # Stage 4: Cleanup/Rollback
    cleanup_agent = types.LlmAgent(
        model="gemini-2.0-flash-exp",
        system_instruction="""
        If state["rollback_needed"], revert changes.
        Clean up temporary resources.
        Report final status.
        """,
        tools=[rollback_tool, cleanup_tool]
    )

    pipeline = types.SequentialAgent(
        agents=[
            validation_agent,
            transform_agent,
            storage_agent,
            cleanup_agent
        ]
    )

    return pipeline

Anti-Patterns to Avoid

❌ Monolithic Agent

# BAD: Single agent trying to do everything
giant_agent = types.LlmAgent(
    model="gemini-2.0-flash-exp",
    system_instruction="""
    You are an expert in Python, TypeScript, Rust, DevOps, security,
    databases, networking, machine learning, data science, mobile dev,
    game dev, blockchain... [10 more domains]

    [3000 line prompt trying to cover everything]
    """,
    tools=[100_different_tools]  # Too many!
)

Problem: Impossible to maintain, debug, or improve. Model gets confused.

Solution: Use coordinator pattern with focused specialists.

❌ Circular Agent Dependencies

# BAD: Agent A calls Agent B which calls Agent A
agent_a = types.LlmAgent(
    tools=[types.AgentTool(agent=agent_b)]  # Calls B
)

agent_b = types.LlmAgent(
    tools=[types.AgentTool(agent=agent_a)]  # Calls A - CIRCULAR!
)

Problem: Infinite loops, stack overflow.

Solution: Design clear hierarchies (coordinator → specialists, no loops).

❌ Using WorkflowAgent for Dynamic Tasks

# BAD: Fixed sequence for unpredictable task
workflow = types.SequentialAgent([
    search_agent,  # What if user doesn't need search?
    analyze_agent,
    summarize_agent  # What if there's nothing to summarize?
])

Problem: Wastes resources on unnecessary steps.

Solution: Use LlmAgent for dynamic tasks, WorkflowAgent only for predictable flows.

❌ Overstuffing Session State

# BAD: Storing large data in session state
context.state["entire_dataset"] = load_1gb_file()  # Too big!
context.state["all_search_results"] = [...]  # Thousands of items

Problem: Performance degradation, memory issues.

Solution: Store references (IDs, paths) in session state, not full data.

When to Use This Skill

Activate this skill when:

• Designing multi-agent systems
• Choosing between LlmAgent and WorkflowAgent
• Implementing coordinator/dispatcher patterns
• Creating hierarchical agent teams
• Optimizing agent communication
• Debugging multi-agent workflows

Integration Points

This skill depends on:

• adk-fundamentals: Single-agent basics
• agentient-python-core/async-patterns: Async coordination

This skill enables:

• Complex production agent systems
• Modular, maintainable agent architectures

Related Resources

For deeper understanding:

• Google ADK Multi-Agent Blog: https://cloud.google.com/blog/products/ai-machine-learning/build-multi-agentic-systems-using-google-adk
• Agent Patterns Guide: https://medium.com/google-cloud/agent-patterns-with-adk-1-agent-5-ways-58bff801c2d6
• ADK Agents Documentation: https://google.github.io/adk-docs/agents/
• Agent-as-Tool Pattern: https://medium.com/@shins777/how-to-use-adk-tools-part-1-function-tools-and-built-in-tools-6e6100f9dfaa