Skip to main content

Module 35: Deploying to Agent Engine

Theory​

From Local Development to Production-Grade Deployment​

Deploying an agent involves packaging its code and running it on a scalable, reliable, and managed platform. Google Cloud's Agent Engine is a fully managed service in Vertex AI designed for this, handling server management, scaling, and security.

Two Paths to Deployment​

  1. Accelerated Deployment (Recommended Best Practice):

    • Method: Use the Agent Starter Pack (ASP) and its command-line interface (uvx agent-starter-pack). This tool applies a production-ready template to your existing ADK project, adding all the necessary files for deployment.
    • Key Features: It includes pre-built CI/CD pipelines using Cloud Build, Infrastructure as Code (IaC) with Terraform, and a simple Makefile for a seamless deployment process.
    • Best For: All new projects. It establishes best practices for security, reliability, and maintainability from the start.

  2. Standard Deployment (Manual):

    • Method: Write a custom Python script using the Vertex AI SDK to package and deploy your agent.
    • Best For: Learning the underlying mechanics of deployment, modifying an existing deployment, or for projects where the starter pack's structure is not a good fit.

This module's lab will guide you through both methods.

The Accelerated Deployment Workflow​

The Agent Starter Pack (ASP) streamlines the path to production into a few simple commands:

  1. Enhance (uvx agent-starter-pack enhance): This is a one-time command that analyzes your existing ADK agent and adds the necessary deployment files to it. This includes the Makefile, Terraform configurations for infrastructure, and Cloud Build YAML files for the CI/CD pipeline.

  2. Deploy (make backend): This command is the main entry point for deployment. It performs two key stages:

    • Provision: It runs a Cloud Build pipeline that uses Terraform to create all the necessary cloud infrastructure for your agent (Artifact Registry, Agent Engine service, service accounts, etc.).
    • Upload & Execute: It packages your agent code, uploads it to Agent Engine, and starts the agent.

Key Takeaways​

  • Google Cloud's Agent Engine is a fully managed service in Vertex AI for deploying, scaling, and securing ADK agents.
  • The Accelerated Deployment method, using the Agent Starter Pack (ASP), is the recommended best practice for new projects, providing a pre-configured CI/CD pipeline and Infrastructure as Code.
  • The Standard Deployment method involves writing a custom Python script with the Vertex AI SDK, which is useful for understanding the underlying mechanics.
  • The Accelerated workflow follows a simple process: Enhance (adds deployment files to your project) and Deploy (provisions infrastructure and uploads your code).
  • Agent Engine and Complex Clients: The architecture of Agent Engine (a managed, scalable backend) simplifies the development of complex clients (like web or mobile apps). It handles concurrency and horizontal scaling automatically, provides a stable API for clients written in any language to connect to, and cleanly separates the UI/session management logic on the client from the heavy lifting of agent orchestration and AI logic on the backend.
  • Advantages of Accelerated Deployment: The Accelerated Deployment method is superior for production as it automates the creation of infrastructure as code (IaC) with Terraform and establishes a simple, repeatable deployment process. This provides a reproducible, auditable, and consistent deployment, incorporating best practices from the start, whereas the Standard Deployment is manual and better suited for learning the underlying SDK mechanics.