Google Agent Development Kit (ADK)

Google's official Python framework for building, evaluating, and deploying AI agents. Released alongside the protocols/a2a protocol in April 2025. By November 2025, it reached v1.18.0 with active ongoing development. Designed to complement — not replace — existing frameworks: LangGraph agents can be wrapped as A2A agents and interoperate with ADK agents.

[Source: google/adk-python GitHub, Google Developers Blog, 2026-05-03]

Key Facts

• GitHub: google/adk-python
• Docs: https://google.github.io/adk-docs
• PyPI: pip install google-adk
• v1.0.0: first stable/GA release — all core interfaces moved to async
• v1.18.0: current release as of November 5, 2025
• A2A native: the reference SDK for building agents that speak the protocols/a2a protocol
• Deploy to: Vertex AI Agent Engine (via google-cloud-aiplatform >= 1.95.0)
• Default model: Gemini (any Gemini model via Vertex AI or Google AI Studio)

Core Pattern

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

root_agent = LlmAgent(
    model="gemini-2.0-flash",
    name="research_agent",
    instruction="You are a helpful research assistant. Use google_search to find current information.",
    tools=[google_search]
)

Agents are Python objects. Tools are Python functions decorated with @tool or built-in tools. The ADK handles the ReAct loop automatically.

Architecture

Agents

• LlmAgent: the standard agent — wraps a Gemini model, tools, and instructions
• SequentialAgent: runs sub-agents in sequence, passing state forward
• ParallelAgent: runs sub-agents concurrently, merges results
• LoopAgent: repeats a sub-agent until a termination condition is met

Tools

from google.adk.tools import FunctionTool

def get_stock_price(ticker: str) -> dict:
    """Fetch the current stock price for a ticker symbol."""
    # ... API call
    return {"ticker": ticker, "price": 182.45, "currency": "USD"}

price_tool = FunctionTool(get_stock_price)

Type hints and docstrings generate the tool schema. The Gemini model sees the schema and decides when to call the tool.

Sessions and Memory

ADK v1.0.0 made all session/artifact/memory service interfaces async:

from google.adk.sessions import InMemorySessionService

session_service = InMemorySessionService()
session = await session_service.create_session(
    app_name="my_agent",
    user_id="user-123"
)

For production, use VertexAiSessionService to persist sessions in Vertex AI.

Multi-Agent with A2A

ADK's primary multi-agent pattern uses the protocols/a2a protocol. Agents expose themselves as A2A services; other agents discover them via Agent Cards at /.well-known/agent.json.

from google.adk.a2a import A2AServer

# Make this agent available as an A2A service
server = A2AServer(agent=research_agent, port=8080)
await server.start()

A root agent can delegate to remote A2A agents:

from google.adk.a2a import A2AClient

remote_agent = A2AClient(url="http://research-agent:8080")
root_agent = LlmAgent(
    model="gemini-2.0-flash",
    name="orchestrator",
    tools=[remote_agent.as_tool()]
)

This lets you build networks of specialised agents built on different frameworks (LangGraph, CrewAI, AutoGen) as long as they implement A2A.

Deployment: Vertex AI Agent Engine

import vertexai
from vertexai.preview import reasoning_engines

vertexai.init(project="my-project", location="us-central1")

app = reasoning_engines.ReasoningEngine.create(
    reasoning_engines.AdkApp(agent=root_agent),
    requirements=["google-adk>=1.0.0"],
)

# Query the deployed agent
response = app.query(input="What's the latest on quantum computing?")

Agent Engine provides managed hosting, session persistence, logging, and horizontal scaling.

vs LangGraph

Connections

• protocols/a2a — ADK is the reference implementation of the A2A protocol
• agents/langgraph — primary competitor; LangGraph agents can be wrapped as A2A agents for ADK interop
• agents/strands-agents-sdk — AWS equivalent; also A2A compatible
• apis/google-ai — Gemini models that power ADK agents
• agents/multi-agent-patterns — Supervisor, parallel fan-out, and sequential patterns ADK implements
• infra/deployment — Vertex AI Agent Engine for managed ADK deployment

Open Questions

• When does ADK beat LangGraph? (ADK wins when you need A2A interop or are all-in on Google Cloud; LangGraph wins when you need fine-grained graph control or HITL interrupts)
• How does ADK Python 2.0 Beta (workflows + agent teams) change the architecture? [unverified — monitor release notes]
• Does ADK support non-Gemini models for production workloads without LiteLLM?