Google Adk

Sessions

• InMemorySessionService: For development/testing
• DatabaseSessionService: For production persistence

Project Structure

adk_agents/
├── agent.py              # Root agent definition
├── _agents/              # Individual agent implementations
│   └── my_agent.py
├── _workflows/           # Workflow orchestrations
│   └── my_workflow.py
├── tools/                # Custom tool implementations
│   └── my_tool.py
├── _state.py             # State key definitions
├── _schemas.py           # Pydantic schemas
└── .adk/                 # ADK config & eval history

Common Patterns

Creating an LlmAgent

from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool

agent = LlmAgent(
    name="my_agent",
    model="gemini-2.0-flash",
    instruction="You are a helpful assistant.",
    tools=[FunctionTool(my_function)],
)

Creating a Sequential Workflow

from google.adk.agents import SequentialAgent

workflow = SequentialAgent(
    name="my_workflow",
    sub_agents=[agent1, agent2, agent3],
)

Using MCP Tools

from google.adk.tools.mcp_tool.mcp_toolset import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StdioConnectionParams

toolset = McpToolset(
    connection_params=StdioConnectionParams(
        server_params=StdioServerParameters(
            command="python",
            args=["-m", "my_mcp_server"],
        )
    ),
    tool_filter=["my_tool"],
)

CLI Commands

Deployment to Vertex AI

# Single command deployment
adk deploy --project=my-project --region=us-central1

# Or via gcloud
gcloud ai agent-engines create my-agent \
  --project=my-project \
  --location=us-central1

Best Practices

1. State Management: Define all state keys in a central _state.py
2. Schema Definitions: Use Pydantic models in _schemas.py
3. Tool Design: Keep tools focused and composable
4. Testing: Use adk eval for systematic agent evaluation
5. Error Handling: Implement graceful fallbacks in agents

Resources

• ADK Documentation
• ADK Python GitHub
• ADK Samples
• Vertex AI Agent Engine