Understanding A2A: How AI Agents Talk to Each Other, Where Humans Stay in Control, and How to Build It with Google ADK

MCP taught AI agents how to use tools. The Agent2Agent (A2A) protocol teaches them how to work with each other. Created by Google and now an open standard under the Linux Foundation, A2A lets agents built by different teams, frameworks and vendors discover each other, delegate tasks and return results, without sharing their code or internals.

This talk starts with the fundamentals on what A2A is, why multi-agent systems need a shared protocol, and how it works under the hood. We'll cover Agent Cards (how an agent advertises its skills), tasks and messages (how work is handed off and tracked), and how A2A complements MCP rather than competing with it.

Next, we'll explore Google's Agent Development Kit (ADK), the open-source framework for building AI agents with Gemini, and see how few lines it takes to publish an agent over A2A or call a remote one. A demo brings it together - a coordinator agent answers an everyday question by delegating parts of it to specialist agents over A2A. We'll inspect a real Agent Card, trace the messages between agents, and walk through the code behind each step.