Prototype ADK Agents with Gemini CLI¶

This demonstration guide walks through the process of rapidly prototyping an ADK (Agent Development Kit) agent using the Gemini CLI. You will leverage the Gemini CLI's planning and implementation capabilities to quickly scaffold and refine a functional customer support agent. This "vibe prototyping" approach allows for fast iteration and development, showcasing how AI-assisted tools can accelerate the creation of complex agents and services. By the end of this demo section, you will have a working ADK agent capable of looking up and summarizing in-memory ticket data, all built and debugged primarily through natural language commands in the Gemini CLI.

Key Files¶

GEMINI.md: Provides instructions on how to use the reference documentation to build ADK agents.
AGENTS.txt: Contains detailed information and best practices for agent development.

AGENTS.txt file is imported into the GEMINI.md and will be included in the Gemini CLI session’s context when you interact with the Gemini CLI.

This feature facilitates the decomposition of large GEMINI.md files into smaller, more manageable modules that can be seamlessly reused across varied contexts. The import processor supports both relative and absolute paths, incorporating robust safety mechanisms to avoid circular imports and ensure secure file access.

Custom Commands¶

These custom commands help you in both planning and implementing ADK agents.

/adk-plan:new - this command leverages provided documentation (GEMINI.md, AGENTS.md) to generate an implementation plan, ensuring adherence to best practices.
/adk-plan:impl - this command uses the implementation plan to generate the necessary Python files, reducing boilerplate and focusing on ADK requirements.

Start building the agent¶

Change the working directory:

git clone --filter=blob:none --no-checkout https://github.com/GoogleCloudPlatform/cloud-solutions
cd cloud-solutions
git sparse-checkout set --cone projects/build-with-gemini-demo/prototype-adk-agent-with-gemini-cli
git checkout
cd projects/build-with-gemini-demo/prototype-adk-agent-with-gemini-cli

Rename .env.sample to .env file and update with your Google Cloud project information:

GOOGLE_GENAI_USE_VERTEXAI=TRUE
GOOGLE_CLOUD_PROJECT="<ENTER_YOUR_PROJECT_ID>"
GOOGLE_CLOUD_LOCATION="<ENTER_YOUR_PROJECT_LOCATION>" # e.g. us-central1

Acquire new user credentials:
```
gcloud auth application-default login
```
Run Gemini CLI:
```
gemini
```

Send prompt to create implementation plan:

/adk-plan:new Build a customer support ADK agent that allows users to look
up the full details of any ticket using its ID and also provide the ability
to return a summary for any selected ticket. For summary requests return
ticket description. Generate 20 sample tickets (each with an integer based
ID, title, and description) and use them as an in-memory db.

Review generated plan and request implementation. You can find an example plan in the projects/build-with-gemini-demo/prototype-adk-agent-sample/plans/customer_support_agent.md file.

Send prompt to implement the plan:
```
/adk-plan:impl implement the plan and generate a requirements.txt
file for more reproducible builds.
```
Review and approve tools and suggested code changes.

The prototype-adk-agent-sample folder contains sample implementation plan and customer support agent built with the prompts above.
Exit from Gemini CLI and run commands to setup virtual environment and install required dependencies for ADK agent:
```
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```
Start ADK Web Server to test the agent:
```
adk web .
```
Click on the link to open application in the browser.

In the Cloud Shell environment, select Web Preview from the menu and change port(eg. 8000) and preview the application.
Testing the Agent

After you start the application and select the agent from the dropdown in the top left corner. Send a greeting to the agent and ask how they can help. This will prompt the agent to describe available operations.

Sample queries to try:
```
lookup ticket # 10
```
```
summarize ticket # 5
```

You can find an example implementation of this agent in the projects/build-with-gemini-demo/prototype-adk-agent-sample directory.