Skip to content

43. Agent-to-Agent Protocol (A2A)

Mini-Project: A2A: Full Client + Server Round-Trip

A complete A2A implementation where a client discovers a server agent's capabilities via its Agent Card, submits tasks using JSON-RPC, and receives structured artifact results — all simulated in pure Python.

View on GitHub

What It Is

The Agent-to-Agent Protocol (A2A) is an open protocol created by Google that enables AI agents built by different vendors to communicate and collaborate. While MCP connects models to tools, A2A focuses on agent-to-agent communication -- allowing agents to discover each other's capabilities, negotiate tasks, and exchange results. A2A defines Agent Cards (capability descriptions), task lifecycle management, and streaming updates.

Who Created It / Governed By

Created by Google (April 2025). Open-source specification with broad industry support from over 50 companies including Salesforce, SAP, and Atlassian.

How It Works

sequenceDiagram
    participant ClientAgent as Client Agent
    participant Card as Agent Card (/.well-known/agent.json)
    participant ServerAgent as Server Agent

    ClientAgent->>Card: GET /.well-known/agent.json
    Card-->>ClientAgent: Agent capabilities, skills, auth
    ClientAgent->>ServerAgent: POST /tasks/send (JSON-RPC)
    ServerAgent-->>ClientAgent: Task accepted + ID
    ServerAgent-->>ClientAgent: Status updates (SSE stream)
    ServerAgent-->>ClientAgent: Final result (artifact)

Key Features

Feature Description
Agent Cards JSON discovery documents at /.well-known/agent.json
Task Model Structured task lifecycle (submitted, working, completed, failed)
Streaming Server-Sent Events for real-time progress updates
Multi-Turn Supports back-and-forth task refinement
Artifacts Structured output objects (text, files, data)
Push Notifications Server can notify client of async completions