Welcome to the a2a_samples repository! This project contains multiple implementations of AI agents using Google's A2A (Agent-to-Agent) protocol.

These implementations demonstrate how to:

• Set up and run A2A-compliant servers and clients
• Use discovery endpoints and standardized task formats
• Integrate with Google's ADK (Agent Development Kit)

We plan to continuously add more versions to illustrate various approaches and frameworks.

a2a_samples/
├── version_1_simple/        # Basic implementation using Flask
├── version_2_adk_agent/     # Advanced agent built using Google ADK + Gemini
├── version_3_multi_agent/   # Multi-agent orchestration example
├── version_4_multi_agent_mcp/ # Distributed multi-agent with A2A + MCP integration
├── version_4p01_with_vision_agent/ # Adds Gemini-based VisionAgent to version_4 architecture
├── version_5_a2a_sdk/       # Minimal A2A PYTHON SDK setup with streaming, LangChain + Gemini
├── version_6_docker_vision_agent/ # Dockerized Gemini Vision Agent deployable to Google Cloud

A beginner-friendly demo that uses Flask to create a basic A2A server agent (TellTimeAgent) and a client that:

• Fetches the agent card from the /.well-known/agent.json endpoint
• Sends a simple task to /tasks/send
• Receives a time-based response

﹖→ Explore the folder

A more advanced version using Google's ADK (Agent Development Kit) to implement a fully functional Gemini-powered A2A agent.

• Integrates memory, sessions, and artifacts
• Uses ADK runners and LLM agents
• Structured with reusable components and aligned with A2A's JSON-RPC model

﹖→ Explore the folder

A multi-agent orchestration example demonstrating:

• Dynamic discovery of TellTimeAgent and GreetingAgent via a registry
• An OrchestratorAgent that routes user queries to the appropriate child agent
• A CLI (cmd.py) for seamless end-to-end testing

﹖→ Explore the folder

A cutting-edge integration combining Google’s A2A protocol with Anthropic’s MCP (Model Context Protocol):

• A2A Protocol – Agents discover each other via JSON-RPC and call one another’s skills.
• MCP Integration – Dynamically discover and load external MCP servers, exposing each tool as a callable function.
• Orchestrator Agent – A central LLM-powered router that decides whether to delegate to a child agent or invoke an MCP tool.
• Modular & Extensible – Simply update a registry or config to add new A2A agents or MCP servers.

﹖→ Explore the folder

An enhanced version of version_4_multi_agent_mcp that adds support for a Gemini-based VisionAgent:

• Accepts image input via local file path or URL
• Uses Google ADK's multimodal capabilities
• Integrates cleanly into the existing A2A and MCP architecture

﹖→ Explore the folder

An educational minimal setup using the official A2A Python SDK with LangChain + Gemini.

• One agent (TellTimeAgent) with time tool
• One async client
• Streaming + multi-turn + structured response support
• Ideal starting point for learning A2A SDK fundamentals

﹖→ Explore the folder

A production-style deployment of a Gemini-based VisionAgent in a Docker container:

• Fully Dockerized architecture for isolated builds
• Local testing using curl to hit the vision agent endpoint
• One-line deployment to Google Cloud Run
• Supports volume-mounted images + container-path resolution

﹖→ Explore the folder

Each version contains its own README.md file with detailed instructions on:

• Setting up Python environments
• Installing dependencies
• Running the server and client

Make sure to check the respective version folder before you begin!

• Add streaming support via SSE and tasks/sendSubscribe
• Add push notification samples
• Add more ADK agent variants with different skills

Stay tuned and ⭐ star the repo if you find it useful!

This repository is licensed under the GNU General Public License v3.0. See the LICENSE file for full details.

Happy building with A2A! 🛠