DEV Community

Cover image for Building AI Agents with the Python Agent Development Kit (ADK) — 2026 Edition (v2)
Google Developer Experts profile image xbill
xbill for Google Developer Experts

Posted on

Building AI Agents with the Python Agent Development Kit (ADK) — 2026 Edition (v2)

#python #ai #googlecloud #gemini

Summary of 2026 Updates (v2 — Python ADK 2.0 Migration)

This updated edition incorporates several key architectural improvements, production hardening, and dependency upgrades over the initial release:

  • Python ADK 2.0 Core Upgrade: Upgraded to google-adk>=2.0.0 (google.adk, GA released at adk.dev/2.0/), introducing google.adk.Context for unified tool context management, session state isolation, and callback execution pipelines.
  • Google Cloud AI Platform Integration: Native authentication auto-detection for Vertex AI (GOOGLE_GENAI_USE_VERTEXAI=TRUE) and Google AI Studio Gemini API Key fallback.
  • Python Toolchain & Environment: Modern Python runtime support (3.10+ up to 3.13) with structured dependency management via requirements.txt.
  • Code & Declarative Agent Support: Native support for Python code-defined agents (agent.py exposing root_agent) as well as YAML-configured agents (root_agent.yaml).
  • Unified Execution Context: Tools accept tool_context: Optional[Context] = None to inspect session metadata, user identity, and execution state.
  • Unified CLI Tooling: Full compatibility with standard ADK 2.x CLI commands (adk run, adk web, adk api_server, adk deploy cloud_run).
  • Web UI & Remote VM Server: Built-in developer Web UI support for both local development (adk web) and remote cloud instances (adk web --host=0.0.0.0).
  • FastAPI API Server Mode: Server mode via adk api_server for embedding agent endpoints into microservice architectures.
  • Serverless Cloud Run Deployment: Streamlined Cloud Run deployment via adk deploy cloud_run featuring embedded Web UI options (--with_ui).

Key Migration Changes from ADK 1.x to ADK 2.x (Python)

If you are migrating an existing codebase from ADK 1.x to ADK 2.x, several key API shifts and architectural additions have been introduced:

Feature / Concept ADK 1.x (Legacy) ADK 2.x (2026 Edition)
Package / Namespace adk / legacy SDK modules google-adk>=2.0.0 (google.adk)
Tool Execution Context Custom kwargs or implicit globals Standard tool_context: Optional[Context] = None
Agent Paradigm Code-only Python classes Dual Code (agent.py) + Declarative YAML (root_agent.yaml)
Workflows & Graphs Custom script loops Native DAG execution via google.adk.Workflow
CLI Tooling Fragmented script execution Unified adk CLI (run, web, api_server, deploy)
Session Persistence Synchronous, in-memory Async persistence (memory://, sqlite://, asyncpg)
Retries & HITL Manual exception catching Framework-managed RetryConfig & HITL pause/resume
Authentication Manual client configuration Auto-detected Vertex AI (GOOGLE_GENAI_USE_VERTEXAI) / Gemini API key

Detailed Breakdown of 1.x -> 2.x Changes:

1. Import Namespace Standardization

In ADK 1.x, imports relied on legacy modules or top-level adk. In ADK 2.x, all official components reside cleanly under the google.adk namespace:

# ADK 1.x (Legacy)
# from adk.agents import Agent

# ADK 2.x (Current)
from google.adk import Context
from google.adk.agents import Agent
Enter fullscreen mode Exit fullscreen mode

2. Unified Tool Execution Context (Context)

In ADK 1.x, tools received only LLM-extracted function arguments, making session isolation or state modification awkward. In ADK 2.x, tools accept an optional tool_context: Optional[Context] = None parameter:

# ADK 2.x Tool Signature
def my_tool(param1: str, tool_context: Optional[Context] = None) -> Dict[str, Any]:
    # Access session state, execution metadata, or user identity
    if tool_context and tool_context.state:
        session_id = tool_context.state.get("session_id")
    ...
Enter fullscreen mode Exit fullscreen mode

3. Graph & Workflow Support (Workflow)

ADK 2.x introduces native Workflow graph definitions alongside standard agents. Developers can construct directed acyclic graphs (DAGs) for multi-step agent pipelines, parallel execution paths, and state transformation nodes.

4. Declarative YAML Agent Specifications

ADK 2.x adds zero-code declarative YAML definitions (root_agent.yaml) for prompt engineers and system configurators, supported side-by-side with programmatic Python agents (agent.py).

5. Unified adk CLI Tooling

ADK 1.x relied on custom scripts or module execution commands. ADK 2.x introduces a unified adk binary supporting standardized operations:

  • adk run <agent_path>: Terminal execution mode
  • adk web: Local graphical chat UI on port 8000
  • adk web --host=0.0.0.0: Cloud VM / Remote binding mode
  • adk api_server <agent_path>: FastAPI microservice server mode
  • adk deploy cloud_run: Production Cloud Run container deployment

6. Asynchronous Session Providers

ADK 2.x features async session adapters supporting persistent backend drivers such as SQLite (sqlite://), PostgreSQL (asyncpg), and in-memory stores (memory://).

7. Framework-Managed Retries & Human-in-the-Loop (HITL)

ADK 2.x provides automatic exception-aware retries via RetryConfig and first-class pause/resume mechanisms for Human-in-the-Loop tool authorization flows.

Native Python Agent Development with the Python ADK

This tutorial provides a comprehensive guide to building, running, and deploying native AI Agents using Python and the official Google Agent Development Kit (ADK) (google-adk).

What Is Python ADK 2.x?

The Python Agent Development Kit (ADK 2.0) (google-adk>=2.0.0) is a code-first framework created by Google for designing, testing, and deploying intelligent AI agents powered by Gemini models.

Key Features:

  • Model-Agnostic & Cloud-Native: Optimized for Gemini models on Google Cloud Vertex AI or Google AI Studio.
  • Unified Context & Tool State: Uses google.adk.Context to pass unified session state, credentials, and Human-in-the-Loop (HITL) context to tools.
  • Graph & Workflow Support: Native support for DAG execution and Workflow graphs in google.adk.
  • Flexible Declarative & Code Paradigms: Define agents programmatically in Python (agent.py) or declaratively via YAML (root_agent.yaml).
  • Built-in Developer CLI & Web UI: Interactive debugging, web chat interface, and API server out of the box.

Official Repositories & Docs:

Key Dependencies & Directory Structure

requirements.txt

The starter stack relies on Python 3.10+ and the latest google-adk release:

# Dev environment
pip
autopep8

# Application dependencies
flask
google-adk>=2.0.0
google-cloud-aiplatform
Enter fullscreen mode Exit fullscreen mode

Project Layout 

adk-hello-world/
├── Makefile                                                    # Make targets for run, test, lint, web, deploy
├── README.md                                                   # Project documentation
├── GEMINI.md                                                   # Developer & agent instructions
├── requirements.txt                                            # Python dependencies
├── init.sh                                                     # Credentials initialization script
├── set_env.sh                                                  # Environment variables script
├── cli.sh / run.sh                                             # Interactive CLI runner scripts
├── Agent1_cli.sh                                               # CLI script for YAML Agent1
├── web.sh / webvm.sh                                           # Web UI launcher scripts
├── api_server.sh                                               # FastAPI server launch script
├── cloudrun.sh                                                 # Cloud Run deployment script
├── tests/
│   └── test_agent.py                                           # Unit tests for agent tools & timezone resolution
└── src/
    └── agents/
        ├── adk_hello_world/                                    # Python-defined agent
        │   ├── agent.py                                        # Agent definition & custom tools
        │   └── requirements.txt
        └── Agent1/                                             # Declarative YAML agent
            └── root_agent.yaml
Enter fullscreen mode Exit fullscreen mode

Checking the Developer Environment

Clone the repository and set up environment credentials:

git clone https://github.com/xbill9/adk-hello-world
cd adk-hello-world
Enter fullscreen mode Exit fullscreen mode

1. Project & Credential Initialization

Run init.sh to configure your Google Cloud Project ID and Gemini API Key:

$ ./init.sh
--- Setting Google Cloud Project ID File ---
Please enter your Google Cloud project ID: my-gcp-project-id
You entered: my-gcp-project-id
Successfully saved project ID.
--- Setting Google Cloud Gemini Key File ---
Please enter your Google Cloud Gemini Key: AIzaSy...
Successfully saved Gemini Key.
--- Setup complete ---
Enter fullscreen mode Exit fullscreen mode

2. Loading Shell Environment Variables

Source set_env.sh to load active environment configurations:

$ source ./set_env.sh
--- Setting Google Cloud Environment Variables ---
Checking gcloud authentication status...
gcloud is authenticated.
Are you using a Gemini API Key? (y/N): n
Configuring for Vertex AI...
Exported PROJECT_ID=my-gcp-project-id
Exported PROJECT_NUMBER=1056842563084
Exported SERVICE_ACCOUNT_NAME=1056842563084-compute@developer.gserviceaccount.com
Exported GOOGLE_CLOUD_PROJECT=my-gcp-project-id
Exported GOOGLE_GENAI_USE_VERTEXAI=TRUE
Exported GOOGLE_CLOUD_LOCATION=us-central1
Exported AGENT_PATH=/home/user/adk-hello-world/src/agents/adk_hello_world
--- Environment setup complete ---
Enter fullscreen mode Exit fullscreen mode

Building Python Agents with ADK 2.x

1. Code-Defined Agent (agent.py)

In ADK 2.x, Python function tools receive function arguments and an optional tool_context: Optional[Context] = None. ADK automatically inspects docstrings and type annotations to generate the tool's OpenAPI schema for Gemini.

Here is src/agents/adk_hello_world/agent.py:

import datetime
from typing import Any
from zoneinfo import ZoneInfo, ZoneInfoNotFoundError

from google.adk.agents import Agent

CITY_TIMEZONE_MAP = {
    # New York region & aliases
    "new york": "America/New_York",
    "new york city": "America/New_York",
    "nyc": "America/New_York",
    "washington dc": "America/New_York",
    "boston": "America/New_York",

    # US Central & Pacific
    "chicago": "America/Chicago",
    "los angeles": "America/Los_Angeles",

    # Europe & Asia
    "london": "Europe/London",
    "paris": "Europe/Paris",
    "tokyo": "Asia/Tokyo",
    "beijing": "Asia/Shanghai",
    "sydney": "Australia/Sydney",
}


def resolve_timezone(city: str) -> str | None:
    """Helper function to resolve a city name or alias to an IANA timezone string."""
    city_clean = city.strip().replace(".", "")
    city_lower = city_clean.lower()
    if city_lower in CITY_TIMEZONE_MAP:
        return CITY_TIMEZONE_MAP[city_lower]

    try:
        ZoneInfo(city)
        return city
    except ZoneInfoNotFoundError:
        pass

    formatted_city = city_clean.title().replace(" ", "_")
    for region in ["America", "Europe", "Asia", "Australia", "Africa", "Pacific"]:
        candidate = f"{region}/{formatted_city}"
        try:
            ZoneInfo(candidate)
            return candidate
        except ZoneInfoNotFoundError:
            pass

    return None


def get_weather(city: str) -> dict[str, Any]:
    """Retrieves the current weather report for a specified city.

    Args:
        city (str): The name of the city for which to retrieve the weather report.
    Returns:
        A structured error explaining that live weather is not configured.
    """
    return {
        "status": "error",
        "error_message": (
            f"Live weather information for '{city}' is not available because "
            "this demo has no weather data provider configured."
        ),
    }


def get_current_time(city: str) -> dict[str, Any]:
    """Returns the current time in a specified city.

    Args:
        city (str): The name of the city for which to retrieve the current time.
    Returns:
        A structured time report or error message.
    """
    tz_identifier = resolve_timezone(city)

    if not tz_identifier:
        return {
            "status": "error",
            "error_message": f"Sorry, I don't have timezone information for {city}.",
        }

    tz = ZoneInfo(tz_identifier)
    now = datetime.datetime.now(tz)
    report = f'The current time in {city} is {now.strftime("%Y-%m-%d %H:%M:%S %Z%z")}'
    return {"status": "success", "report": report}


root_agent = Agent(
    name="weather_time_agent",
    model="gemini-2.5-flash",
    description=(
        "Agent to answer questions about the time and weather in a city."
    ),
    instruction=(
        "You are a helpful agent who can answer questions about the current "
        "time in supported cities. The weather tool does not have a live data "
        "provider; clearly tell users when weather data is unavailable."
    ),
    tools=[get_weather, get_current_time],
)
Enter fullscreen mode Exit fullscreen mode

2. Declarative YAML Agent (root_agent.yaml)

ADK 2.x also supports declarative agent definitions in YAML, such as src/agents/Agent1/root_agent.yaml:

name: Agent1
model: gemini-2.5-flash
agent_class: LlmAgent
instruction: You are the root agent that coordinates other agents.
sub_agents: []
tools:
  - name: google_search
Enter fullscreen mode Exit fullscreen mode

Running the ADK Agent Locally

Option A: Interactive Terminal CLI (cli.sh / run.sh)

Launch an interactive terminal session with your code-defined agent using cli.sh or make run:

$ source ./cli.sh
/home/user/adk-hello-world/src/agents/adk_hello_world
adk run .

User -> What time is it in Tokyo?

Agent -> The current time in Tokyo is 2026-07-27 04:12:48 JST+0900.

User -> What is the weather in New York?

Agent -> Live weather information for 'New York' is not available because this demo has no weather data provider configured.
Enter fullscreen mode Exit fullscreen mode

To run the YAML-configured agent Agent1:

$ source ./Agent1_cli.sh
/home/user/adk-hello-world/src/agents/Agent1
adk run .
Enter fullscreen mode Exit fullscreen mode

Option B: Local Web UI (web.sh / webvm.sh)

Launch the built-in visual web interface:

$ source ./web.sh
/home/user/adk-hello-world/src/agents
adk web
Enter fullscreen mode Exit fullscreen mode

Open http://localhost:8000 in your browser to access the graphical chat UI, inspect tool call payloads, and view real-time agent execution traces.

For remote Cloud VMs or headless setups, use webvm.sh to bind to all network interfaces:

$ source ./webvm.sh
Running ADK from Cloud VM
/home/user/adk-hello-world/src/agents
adk web --host=0.0.0.0
Enter fullscreen mode Exit fullscreen mode

Option C: FastAPI API Server (api_server.sh)

Start an API server endpoint to expose your agent over HTTP:

$ source ./api_server.sh
setting API Server Mode
/home/user/adk-hello-world/src/agents/adk_hello_world
adk api_server .
Enter fullscreen mode Exit fullscreen mode

Google Cloud Run Deployment (cloudrun.sh)

Deploy your Python ADK agent directly to Google Cloud Run in a single command using cloudrun.sh:

$ source ./cloudrun.sh
adk deploy cloud_run \
  --project=my-gcp-project-id \
  --region=us-central1 \
  --service_name=hello-world-agent-service \
  --app_name=hello-world-agent-app \
  --with_ui \
  /home/user/adk-hello-world/src/agents/adk_hello_world
Enter fullscreen mode Exit fullscreen mode

The CLI builds the container image and provisions a fully-managed serverless endpoint on Cloud Run equipped with the web UI enabled (--with_ui).

Summary & Next Steps

In this updated 2026 edition:

  1. Migrated agent definitions to Python ADK 2.x (google-adk>=2.0.0) running on Python 3.10+.
  2. Updated custom tools to leverage google.adk.Context for unified session state and execution pipelines.
  3. Highlighted key 1.x to 2.x migration changes, including import namespaces, async session stores, Workflow DAG execution, declarative YAML definitions, and RetryConfig / HITL tools.
  4. Demonstrated both code-defined agents (agent.py) and declarative YAML agents (root_agent.yaml).
  5. Validated local execution across terminal CLI (adk run), Web UI (adk web), and FastAPI server (adk api_server).
  6. Streamlined serverless deployment to Google Cloud Run with adk deploy cloud_run.

To explore more advanced capabilities such as multi-agent orchestration, custom callback middleware, and async persistence adapters, visit the official documentation at adk.dev/2.0/.

Top comments (0)