Build an AI agent with Gemini CLI and Agent Development Kit

Agentverse Champions: Developer – Shadow Blade Mission Summary

Key Concepts:

• Gemini CLI: A natural language interface for interacting with Gemini AI models directly from the terminal, enabling “vibe coding.”
• Vibe Coding: Transforming intent into functional applications using natural language prompts.
• MCP (Model Context Protocol): A specialized portal allowing AI agents to interact with external tools and services.
• Context Engineering: Providing instructions to an LLM at different levels (system, short-term memory, long-term memory) to optimize output and enforce guardrails.
• ADK (Agent Development Kit): A toolkit for building, evaluating, and deploying AI agents.
• CI/CD Pipeline: Automated process for testing, building, and deploying code.
• Cloud Run: Google’s serverless platform for deploying containerized applications.
• Makaker’s Mark: A developer’s signature style, represented in this case by a personal profile website.

1. Introduction & Mission Overview

The Agentverse Champions series focuses on assembling a team of specialists to build and deploy a powerful, secure, and intelligent application. This segment focuses on the developer’s role – the “Shadow Blade” – and their mission to master the Gemini CLI and build a personal profile website. The ultimate goal is to demonstrate the power of “vibe coding” and agent-based development. The speaker emphasizes that the video provides a high-level overview, and a detailed lab is available in the description for full completion.

2. Mastering the Gemini CLI – The Primary Weapon

The Gemini CLI is presented as more than just a tool; it’s an extension of the developer’s will, enabling interaction with Gemini AI models through natural language prompts. The “yellow flag” is introduced as a way to bypass confirmation prompts, accelerating development in non-production environments, but with a caution to use it carefully. The initial task is to create a basic personal profile website using a single command: instructing Gemini to create index.html and styles.css files. A simple Python web server is then launched to preview the website locally, accessible via web preview on port 8000. The speaker encourages experimentation with the website’s theme and colors through further natural language instructions.

3. Integrating with External Tools via MCP Servers

To avoid losing progress and manage inevitable AI misbehavior, version control is introduced using a local Git T server. This server is connected to the Gemini CLI through an MCP server. MCP servers are described as specialized portals enabling AI agents to interact with external tools. The Git T server and its corresponding MCP server are spun up, and the Gemini CLI is configured to recognize the Git arsenal by updating the settings.json file. The CLI is then used to create a repository ("shadowblade profile") and commit the website files to it.

4. Enhancing the Website with AI-Generated Imagery

The website is identified as lacking a profile picture. An issue is logged in the Git repository using the Gemini CLI, demonstrating its ability to function as a single interface for coding and issue tracking. An image creation model from Vert.Ex AI is then integrated via another MCP server. A "sacred bucket" is created to store the generated image. The Gemini CLI is used to generate a pixel art portrait of a “shadow blade” and then modify the index.html file to display the new image. The updated website is previewed, and the changes are committed to the Git repository, closing the issue.

5. Implementing Guardrails with Context Engineering

The speaker highlights the potential for LLMs to produce unpredictable results, emphasizing the need for “guardrails” to ensure discipline and quality. This is achieved through “context engineering” – providing instructions at different levels: system level (instructions/prompts), short-term memory (conversation history), and long-term memory (policies, style guides). The agentesign.mmd file (retrieved from the agentverse) is described as a blueprint for clarity and technical implementation. A gemini.md file is created in the project root, containing Python best practices to serve as a persistent project-level instruction for the agent.

6. Building and Evaluating the Agent

An agent is created using a single command that incorporates the newly defined guardrails. The agent’s core logic is established, and its toolset is defined using MCP servers. The MCPs server.py file is identified as defining the agent’s custom MCP server and exposing tools using the ADK’s MCP tool decorator. Seven new tools are added to the agent’s arsenal using the Gemini CLI.

7. Automating Testing with CI/CD and ADK Eval

A CI pipeline is built to automate testing and building the code. ADK eval is used to run the agent against a “golden dataset” of challenges and synthetic data, verifying its responses. The evaluation is defined by an eval set.json file and a test config.json file. The agent successfully passes the ADK eval test. Piest is then used to further automate the evaluation process, creating a test file and running it against the eval set. The test passes, confirming the agent’s logic and tool selection.

8. Deployment to Cloud Run

A Cloud Build.yaml file is generated using the Gemini CLI, instructing Cloud Build to run the PI test script, build a Docker container if the test passes, and save it to artifact registry. The build process is initiated, and the resulting container is deployed to Cloud Run, making the agent a live public service.

9. Conclusion & Next Steps

The developer’s mission is complete, having successfully vibecoded an application, integrated it with external tools, enforced quality with guardrails, and deployed it to the cloud. The speaker encourages further exploration of the Agentverse and collaboration with other specialists (data engineer, platform engineer, architect). Links to the lab, blog post, and video are provided in the description.

Notable Quotes:

• “Gemini CLI is more than a tool. It’s an extension of your will.”
• “A blade without discipline is a danger.”
• “Vibe coding is all fun and games until the vibes are off.”

Technical Terms:

• LLM (Large Language Model): A type of AI model trained on a massive dataset of text.
• Docker: A platform for developing, shipping, and running applications in containers.
• Artifact Registry: A service for storing and managing container images and other artifacts.
• YAML: A human-readable data serialization language often used for configuration files.
• JSON: A lightweight data-interchange format.