Build a multi-agent system: A2A & Agent Registry

Github repo → https://goo.gle/multiagent When you have 100 AI agents across your org, how do they find and talk to each other? This is how: the Agent2Agent (A2A) Protocol and Agent Registry. In this episode of Hands On AI, we build a production-ready multi-agent system. We start with the Agent2Agent (A2A) Protocol, think "HTTP for AI agents", then solve agent sprawl with Agent Registry, so your agents discover each other instead of hard-coding URLs. We build it live with a dog walker agent and a trip planner agent on Google's Agent Development Kit (ADK), then orchestrate them together. What you'll learn: * What the A2A protocol is, and when to use it for multi-agent systems. * Agent cards — an agent's "LinkedIn profile" for discovery. * Sync vs async agent communication (polling vs streaming). * Wrapping an ADK agent as an A2A remote agent with to_a2a. * The agent sprawl problem — and how Agent Registry solves it (reusability, one way to connect, governance). * Registering agents via gcloud and the Cloud console. * Using Agent Registry's MCP server to discover agents + tools dynamically — no more hard-coded URLs. * Bringing in third-party agents (LangGraph, CrewAI) and MCP servers. Chapters: 00:00 - Intro: Agent2Agent (A2A) Protocol & Agent Registry 03:06 - What Is Agent2Agent (A2A) Protocol? (HTTP for AI agents) 04:29 - When & why to use A2A 06:31 - Agent cards: An agent's LinkedIn profile 09:34 - Sync vs async communication 11:20 - Demo: building a dog walker agent (ADK) 15:21 - Wrapping an agent as an A2A remote agent 17:20 - Connecting two agents with A2A 19:27 - The problem: Agent sprawl 20:05 - What is Agent Registry? 22:35 - Registering agents (gcloud + console) 27:13 - Beyond agents: MCP servers & endpoints 29:18 - Connecting through the registry — No more hard-coded URLs 35:36 - Demo: The orchestrator agent 41:09 - Authentication & wrap up More resources: Google Agent Development Kit (ADK) → https://goo.gle/3PItVzL Agent Runtime → https://goo.gle/4nXD