Java Implementation of Google's A2A Protocol (A2A-Compliant)

This project demonstrates how to build an A2A-compliant multi-agent system using Spring Boot, Ollama, and the Agent-to-Agent (A2A) protocol.

The architecture is based on a Parent Agent that orchestrates multiple Child Agents (e.g., Order Agent, Seller Agent). The parent dynamically discovers child agents via their AgentCard (agent.json) and delegates tasks in a standardized, message-driven way.

Key Features

A2A-Compliant Communication – Agents communicate through the A2A protocol using standardized Message objects and AgentCard descriptors.

Dynamic Agent Discovery – The parent automatically fetches AgentCards from configured child agents, ensuring no hardcoding of capabilities.

LLM-Driven Task Routing – The parent uses a Large Language Model (via Ollama) to decide which agent(s) should handle a given user query.

Chained Orchestration – The parent can call multiple agents in sequence, e.g., fetching order details from the Order Agent and then using the seller ID to fetch additional info from the Seller Agent.

Response Aggregation – The parent aggregates responses from multiple agents into a final, natural-language answer for the user.

Strict Agent Boundaries – Each agent is independent and exposes its own agent.json, making the system modular and extendable.

Use Cases

Case 1: Get B2C Orders → Only the Order Agent is called.

Case 2: Get Seller Info by ID → Only the Seller Agent is called.

Case 3: Get B2C Orders with Seller Info → The Order Agent is called first, then the Seller Agent is called using the extracted seller ID(s).

Flow Diagram

Testing with Prompts