GCP and building AI Agents
Exploring GCP and Building AI Agents: A Hands-On Day with Vertex AI and the Agent Development Kit
Today was one of those days where I fully leaned into the world of Google Cloud Platform (GCP) and got deep into building AI agents. After spending some time exploring Microsoft’s Copilot Agents, I figured it was only fair to give Google a shot—and I’m glad I did.
My goal? Build a multi-agent system with a manager, persistent storage, and integration into a lightweight LLM. This project gave me the perfect opportunity to dive into Google’s Vertex AI Agent Builder, explore LangChain, and experiment with Google’s Agent Development Kit (ADK).
Here’s what stood out—and why it was such an exciting day.
Giving Google a Fair Shot: Multi-Agent Design with GCP
After seeing what Microsoft had to offer, I wanted to explore how GCP could support something similar, but more tailored to my vision: intelligent agents with specific purposes that could collaborate and be tested easily.
Enter: Vertex AI Agent Builder + the ADK + LangChain. This combo gave me the tools to automate workflows while keeping things modular, testable, and organized.
Agent Development Kit (ADK):
Let’s talk about a few things I loved about working with the Agent Development Kit.
Quick Testing with adk web
This might seem small, but it’s actually a huge win. With a single command—adk web—you can spin up a chat-based interface to test your agents live. No extra setup. No complex configurations.
If you’ve ever worked with agents before, you know how refreshing it is to test them without setting up a bunch of glue code. This feature alone saved me time and made development feel way smoother.
Natural Language Agent Definitions Here’s an example of one of the agents I built today:
osloganomalydetectionagent = LlmAgent( name="OsLogAnomalyDetectionAgent", model=GEMINI_MODEL, instruction=""" You are an OS Log Anomaly Detection Agent. Your primary role is to monitor and analyze operating system log files for anomalies, errors, security events, and any deviations from normal system behavior.
When asked to analyze OS log data, you should:
1. Use the 'get_os_log_data' tool to retrieve raw OS log entries.
2. Carefully scan the returned 'logs' data for unusual patterns, critical error messages, suspicious activities (e.g., multiple failed login attempts, unauthorized access attempts, unusual process terminations), or any indicators of system instability or security breaches.
3. Identify and categorize specific anomalies, noting their timestamps, the full log entry (or a relevant snippet), their severity, and suggest potential actions.
4. If no significant anomalies are found, state that the logs appear normal.
""",
description="Monitors and analyzes OS log files for anomalies, errors, and security events, providing detailed reports.",
tools=[get_os_log_data],
output_key="os_log_anomalies"
)
The most powerful thing here? You define the behavior and intelligence of the agent in plain language. It’s readable, flexible, and surprisingly precise—perfect for describing exactly what the agent should do without overcomplicating it.
- Enforced Directory Structure = Instant Organization Here’s a little thing that makes a big difference: ADK requires your project directory name to match your agent’s name.
Why does that matter?
It forces you to be organized. You’ll never lose track of your agent files because everything is exactly where it should be. This kind of structure is a quiet hero—it saves time, reduces bugs, and makes it easier to collaborate later on.
Final Thoughts
Today wasn’t just a “learn some tech” kind of day—it was a build, test, and get inspired kind of day. GCP, Vertex AI, and the ADK made the process of developing and managing intelligent agents actually enjoyable.
Between the simplicity of adk web, the natural language agent setup, and the forced structure that keeps everything tidy, I can confidently say Google’s tools have some serious potential in the multi-agent space.
If you're even remotely curious about creating AI agents that feel intelligent and useful, I highly recommend giving the Agent Development Kit a try. Trust me, it’ll be one of those days where your brain is tired—but in the best way.