Google Agent Development Kit (ADK)

Google Agent Development Kit (ADK) is an open-source Python and Java toolkit that lets developers build, evaluate, and deploy production AI agents using a code-first approach. Released under the Apache 2.0 license, ADK provides a structured framework for creating single agents or complex multi-agent systems without requiring a paid license or vendor lock-in.

ADK reached its Python 1.0.0 stable release in May 2025, and a Java SDK (0.1.0) followed shortly after. The framework is the same one Google uses internally to power agents across Agentspace and other first-party products, which means it receives continuous investment from Google's engineering teams rather than relying solely on open-source contributors.

At its core, ADK treats agents as composable Python or Java classes. Each agent is defined with natural-language instructions, a set of callable tools (including custom functions, API connectors, and Google Search grounding), and optional sub-agents it can delegate work to. This class-based architecture means agents live alongside your existing application code — they're version-controlled, unit-testable, and deployable through standard CI/CD pipelines.

The multi-agent orchestration system is one of ADK's strongest differentiators. Rather than bolting on multi-agent support as an afterthought, ADK ships with first-class primitives: SequentialAgent for step-by-step pipelines, ParallelAgent for concurrent execution, LoopAgent for iterative refinement, and a coordinator pattern for hierarchical delegation. These primitives handle state passing, error propagation, and agent communication out of the box, eliminating the custom glue code that teams typically write when orchestrating multiple agents in frameworks like LangChain or CrewAI.

ADK also includes a built-in evaluation framework — a feature that most competing frameworks lack entirely. Developers define evaluation sets with expected agent trajectories and final responses, then run scored evaluations locally or in CI. This makes it possible to catch regressions when updating prompts, swapping models, or modifying tool configurations before changes reach production.

While ADK is optimized for Google's Gemini models (including Gemini 2.5 Flash and Pro with their 1M-token context windows), the framework is genuinely model-agnostic. Through LiteLLM integration, developers can swap in OpenAI GPT-4o, Anthropic Claude, Meta Llama, Mistral, or any OpenAI-compatible endpoint without changing their agent logic. This flexibility matters for teams that want to benchmark across providers or use different models for different agent roles within the same system.

Deployment options range from local development with a simple Gemini API key (available free from Google AI Studio) to fully managed production hosting on Vertex AI Agent Engine. The Vertex AI path provides autoscaling, load balancing, IAM-based access control, VPC Service Controls, audit logging via Cloud Trace, and enterprise SLAs — all accessible through a single deployment command. Self-hosting on other cloud providers or on-premises is also supported, though the documentation for non-Google deployment paths is less mature.

ADK natively supports bidirectional audio and video streaming through the Gemini Live API, enabling real-time voice agents and video-aware assistants with sub-second latency. This capability positions ADK for use cases like live customer support, accessibility tools, and interactive tutoring that require streaming rather than batch request-response patterns.

The framework includes session state management and memory persistence, allowing agents to maintain context across conversations. Tool definitions support custom Python functions, Google Cloud connectors (BigQuery, Cloud Storage, Cloud Functions), and third-party API integrations. While ADK's integration ecosystem is smaller than LangChain's 700+ connectors, the integrations that exist are well-tested and production-grade.

For teams evaluating ADK against established alternatives, the key trade-offs are clear: ADK offers cleaner APIs, better Google Cloud integration, and built-in evaluation tooling, but has a smaller community, fewer third-party integrations, and less battle-tested production usage outside of Google's own products. The framework is best suited for engineering teams comfortable writing Python or Java who want a structured, testable approach to agent development — especially those already invested in the Google Cloud ecosystem.

Step 1: Install ADK

Install via pip: pip install google-adk. Requires Python 3.9+. For Java, add the google-adk Maven dependency to your project.

Step 2: Get a Gemini API Key

Sign up at Google AI Studio and generate a free API key. Set it as an environment variable: export GOOGLEAPIKEY=your-key-here. No credit card required for the free tier.

Step 3: Create Your First Agent

Create a Python file and define your agent:

python
from google.adk.agents import Agent
my_agent = Agent(
    name="my_assistant",
    model="gemini-2.5-flash",
    instructi"You are a helpful assistant that answers questions clearly and concisely.",
)

Step 4: Add Tools

Give your agent capabilities by defining tool functions:

python
def get_weather(city: str) -> dict:
    """Get the current weather for a city."""
    return {"city": city, "temp": "72°F", "condition": "sunny"}
agent = Agent(
    name="weather_agent",
    model="gemini-2.5-flash",
    instructi"Help users check the weather.",
    tools=[get_weather],
)

Step 5: Run and Test Locally

Use the ADK CLI to start a local development server: adk web launches an interactive UI at localhost:8000 where you can chat with your agent, inspect tool calls, and debug behavior.

Step 6: Evaluate Your Agent

Define evaluation cases to test agent behavior:

python
from google.adk.evaluation import EvalSet
evalset = EvalSet(cases=[{"input": "Weather in NYC?", "expectedtools": ["get_weather"]}])

Run evaluations with adk eval to score trajectories and catch regressions before deploying.

Step 7: Deploy to Production

For managed hosting, deploy to Vertex AI Agent Engine: adk deploy cloud_run or configure Vertex AI Agent Engine through the Google Cloud Console. For self-hosting, package your agent as a standard Python application and deploy to any container runtime.