Building A Multi-Agent System

Building a Multi-Agent System

1. Introduction

Overview

In this lab, you will go beyond simple chatbots and build a distributed multi-agent system.

While a single LLM can answer questions, real-world complexity often requires specialized roles. You don't ask your backend engineer to design the UI, and you don't ask your designer to optimize database queries. Similarly, we can create specialized AI agents that focus on one task and coordinate with each other to solve complex problems.

You will build a Course Creation System consisting of:

Researcher Agent: Using google_search to find up-to-date information.

Judge Agent: Critiquing the research for quality and completeness.

Content Builder Agent: Turning the research into a structured course.

Orchestrator Agent: Managing the workflow and communication between these specialists.

What you'll do

Define a tool-using agent (researcher) that can search the web.

Implement structured output with Pydantic for the judge.

Connect to remote agents using the Agent-to-Agent (A2A) protocol.

Construct a LoopAgent to create a feedback loop between the researcher and judge.

Deploy the multi-agent system to Google Cloud Run.

Architecture & Orchestration Principles

Before we write code, let's understand how these agents work together. We are building a Course Creation Pipeline.

The System Design

Architecture Diagram

Orchestrating with Agents

Standard agents (like the Researcher) do work. Orchestrator Agents (like LoopAgent or SequentialAgent) manage other agents. They don't have their own tools; their "tool" is delegation.

LoopAgent: This acts like a while loop in code. It runs a sequence of agents repeatedly until a condition is met (or max iterations reached). We use this for the Research Loop:

Researcher finds info.

Judge critiques it.

If Judge says "Fail", the EscalationChecker lets the loop continue.

If Judge says "Pass", the EscalationChecker breaks the loop.

SequentialAgent: This acts like a standard script execution. It runs agents one after another. We use this for the High-Level Pipeline:

First, run the Research Loop (until it finishes with good data).

Then, run the Content Builder (to write the course).

By combining these, we create a robust system that can self-correct before generating the final output.