Short Course · 16 lessons + 1 interview questions · ~137 mins
A2A: The Agent2Agent Protocol
An open protocol for building multi-agent AI systems — where agents discover, delegate, and collaborate across six frameworks and three model providers.
How agents advertise themselves via Agent Cards, the eight-state Task lifecycle, JSON-RPC 2.0 transport, Server-Sent Events for streaming, and the authentication model — the full spec, explained through working code.
Six frameworks, one protocol
Build the same loan validation problem six ways: Microsoft Agent Framework, Google ADK, LangGraph with MCP, CrewAI, OpenAI Agents SDK, and a bare-metal agent loop. Each produces a standards-compliant A2A server — interchangeable at the protocol layer.
A production-grade capstone
A multi-agent loan approval pipeline with six orchestrated agents, AI-driven decisioning, human-in-the-loop escalation, a React approval dashboard, and OpenTelemetry distributed tracing across the full pipeline.
Local-first throughout
GitHub Models Phi-4 (free with a GitHub account), Azure AI Foundry Kimi-K2 and Kimi-K2-Thinking (free tier), and Foundry Local Qwen2.5 Coder (runs on-device). Every lesson runs without surprise API costs.
About This Course
Most agents today work in isolation. They call tools through MCP, fetch data, execute tasks. But give three agents — built with different frameworks by different teams — a shared problem, and you are back to writing custom glue for every pair. The integration count grows with N. A2A removes that ceiling: each agent implements the protocol once, and any two agents that speak A2A can discover and work with each other without any bespoke adapter code.
The protocol is precise about what it standardises. An Agent Card at /.well-known/agent-card.json declares the agent's name, capabilities, skills, and authentication scheme. Tasks move through eight defined states — submitted, working, input-required, auth-required, completed, canceled, failed, and rejected. Messages carry typed Parts (text, data, files). Streaming responses arrive over Server-Sent Events. Authentication (OAuth 2.0, mTLS, or API key) is declared in the card, not negotiated per-call. Everything runs over JSON-RPC 2.0.
In this course you build the full picture. A QA agent from scratch using the A2A Python SDK. Then the same loan validation problem solved six ways — Microsoft Agent Framework, Google ADK, LangGraph backed by MCP tool servers, CrewAI, the OpenAI Agents SDK, and a bare-metal loop that shows what every framework automates. Each runs as a separate process on a fixed port. The capstone connects all six: a loan pipeline where AI handles 80% of decisions automatically, humans review the rest through a React dashboard, and every step is traced with OpenTelemetry.
1Understand why A2A exists
Explore the client-server architecture of A2A: what an Agent Card is, how tasks flow through the lifecycle (submitted → working → completed), and why standardizing inter-agent communication matters.
2Build and expose your first A2A agent
Build a QA agent, wrap it in an A2A server using the Python SDK, and create an A2A client from scratch — covering Agent Cards, task lifecycle (submitted → working → completed), and Server-Sent Events streaming.
3Integrate six agentic frameworks
Build agents with Microsoft Agent Framework, Google ADK, LangGraph + MCP, CrewAI, OpenAI Agents SDK, and Claude Agent SDK — each wrapped as an A2A server.
4Combine LangGraph with MCP tools
Create a CodeAgent using LangGraph, connect it to FastMCP tool servers for structured tool access, and run it entirely locally with Foundry Local Qwen2.5 Coder.
5Orchestrate with Microsoft Agent Framework
Build an OrchestratorAgent using Microsoft Agent Framework with Kimi-K2-Thinking that routes tasks to specialist A2A agents based on intent classification.
6Build the capstone multi-agent system
Build the capstone loan approval pipeline — six orchestrated agents with AI-driven decisioning, human-in-the-loop escalation, a React approval dashboard, and OpenTelemetry observability.
7Production patterns
Explore A2A extensions, TLS/OAuth 2.0/mTLS security hardening, OpenTelemetry distributed tracing, and GDPR/HIPAA compliance considerations for enterprise deployment.
Who Should Join?
Prerequisites
What you need before starting
PythonAI AgentsMulti-Agent SystemsGit
AI developers building multi-agent systems or working with agentic workflows. Familiarity with Python 3.11+ and a basic understanding of AI agents is recommended. No cloud accounts required — all model providers offer free tiers or local inference.
🛠
AI / ML Engineers
You build AI-powered products and want a standard protocol to connect multiple agents — across frameworks, teams, and organizational boundaries.
⚙️
Backend & Platform Engineers
You design distributed systems and want to add agent-to-agent communication as a first-class capability in your architecture.
💻
Full-Stack Developers
You are exploring agentic AI and want a structured, code-first introduction to multi-agent orchestration with real, runnable examples.
📊
Technical Leaders
You evaluate emerging technologies and need to understand A2A's design, ecosystem backing, and what production readiness looks like.
Course Outline
16 lessons + 1 interview questions · ~137 mins
Each lesson builds on the previous one — follow them in order for the best experience.
Technologist with 20+ years of engineering experience and an ML/AI practitioner since 2010. Founder of ErgoSum (quantitative & equity research) and LocalM (AI-assisted SDLC). Currently focused on AI Platform Engineering, Agentic AI, and Context Engineering.
