Google ADK — Official AI Agent Dev Kit

What it is

Google ADK (Agent Development Kit) is Google's official open-source Python toolkit for building AI agents. It provides a code-first framework for defining agents with tool use, multi-agent orchestration, evaluation, and deployment capabilities. ADK supports Gemini models natively and integrates with the Google AI ecosystem.

ADK targets Python developers building AI agents who want an officially supported framework from Google. It handles agent definition, tool binding, conversation management, multi-agent delegation, and evaluation in a unified API.

How it saves time or tokens

ADK provides the scaffolding for agent development that you would otherwise build yourself: tool registration, conversation state management, multi-agent routing, and evaluation harnesses. The CLI (adk create, adk run, adk eval) accelerates the development loop. Built-in evaluation tools let you test agent behavior systematically before deployment.

How to use

1. Install ADK: pip install google-adk.
2. Create an agent project: adk create my_agent.
3. Define your agent, add tools, and run it: adk run my_agent.

Example

from google.adk import Agent, Tool

# Define a tool
@Tool
def search_docs(query: str) -> str:
    '''Search documentation for relevant information.'''
    # Your search implementation
    return f'Results for: {query}'

# Create an agent
agent = Agent(
    name='doc_assistant',
    model='gemini-2.0-flash',
    tools=[search_docs],
    instruction='You are a documentation assistant. Use search_docs to find answers.'
)

# Run via CLI
# adk run doc_assistant

# CLI workflow
pip install google-adk
adk create my_agent
adk run my_agent
adk eval my_agent --test-file tests.json

Related on TokRepo

• Agent Tools — AI agent frameworks and orchestration
• Coding AI Tools — Developer tools for AI applications

Common pitfalls

• ADK is optimized for Gemini models. While other models may work through adapters, Gemini gets the best support for tool calling and multi-turn conversations.
• Multi-agent orchestration adds complexity. Start with a single agent and add delegation only when a single agent cannot handle the task scope.
• Evaluation test files require structured input/output pairs. Invest in creating good test cases early to catch agent regressions.