ADK Tutorial

This project contains a collection of agents built with the Google Agent Development Kit (ADK).

Getting Started

This project uses the Google Agent Development Kit (ADK). For more information about the ADK, please refer to the official documentation: https://google.github.io/adk-docs/get-started.

Prerequisites

• Python 3.11 or later
• Git

Cloning the Repository

To clone this repository, run the following command:

git clone https://github.com/ragoler/ADK.git
cd ADK

Installation and Usage

1. Create and activate a virtual environment:

   python -m venv .venv
   source .venv/bin/activate

2. Install the dependencies:

   pip install -r requirements.txt

3. Set up your API keys:

   Some agents may require API keys to function. These are configured in a .env file within the agent's directory. Before running an agent, make sure to configure the corresponding .env file with your API keys.

The Agents

This project includes the following agents, each located in its own directory:

• adk-streaming: A simple agent that uses Google Search to answer questions, demonstrating the streaming capabilities of the ADK.
• all_together_agent: An example based on the "Putting it all together" section of the ADK documentation. It demonstrates two agents: one that uses a tool to find the capital of a country, and another that uses an output schema to provide structured information.
• CodeAgent: A sequential code development pipeline (Python-based). It iterates through writing, reviewing, refactoring, and saving code to a structured directory.
• CodeAgentYaml: A YAML-based implementation of the code development pipeline, demonstrating how to define agents and their orchestration using YAML configuration files.
• basket_ball: A multi-agent system that can provide NBA standings and results.
• multi_tool_agent: An agent that can use multiple tools to answer questions about the weather, time, and famous buildings in New York.
• my_agent: A simple agent that can tell the current time in a specified city.
• weather_team: A multi-agent system that can provide weather forecasts, greet the user, and say goodbye. It also includes a callback to check for inappropriate language.
• CreativeWritingLoop: A writer-reviewer loop with iteration visibility and critique-driven revisions.
• ParallelResearchAgent: A dynamic parallel research pipeline that extracts topics from user input and researches them concurrently.

Running the Agents

You can run any of the agents interactively from the root of the project using the adk run <agent_directory> command.

Examples:

# Run the simple time-telling agent
adk run my_agent

# Run the multi-tool agent
adk run multi_tool_agent

# Run the weather team multi-agent system
adk run weather_team

# Run the sequential code pipeline (Python version)
echo "Write a python function that calculates factorial" | adk run CodeAgent

# Run the sequential code pipeline (YAML version)
echo "Write a python function that sorts a list" | adk run CodeAgentYaml

# Run the creative writing loop
adk run CreativeWritingLoop

# Run the parallel research pipeline
adk run ParallelResearchAgent

Designing and running Agents

You can use the adk web --port 8000 command to visualize and edit an agent.

adk web --port 8000

This will start a web server and open the ADK designer in your browser, where you can inspect the agent's properties, tools, and instructions.

GKE Deployment

You can deploy the ADK web interface to Google Kubernetes Engine (GKE) for hosting.

Prerequisites

• Google Cloud project with billing enabled.
• gcloud CLI installed and authenticated.
• kubectl installed.

Deploying to GKE

Run the provided deployment script. This script will automate the creation of a GKE Autopilot cluster, Artifact Registry repository, IAM permissions, and the Kubernetes deployment itself.

./deploy.sh

Accessing the Web Interface

The service is deployed with a ClusterIP for security. To access it locally:

1. Port-forward the service:

   kubectl port-forward svc/adk-web 8000:8000

2. Open in Browser: Visit http://localhost:8000.

Teardown

To delete all resources created by the deployment script (GKE cluster, AR repo, the GSA, and Kubernetes resources):

./deploy.sh --delete