Multi-agent system for query processing with safety verification and critique built with Google A2A protocol, Google ADK, Llama Prompt Guard 2, Gemma 3 and Gemini 2.0 Flash.

A2A Double Validation

Multi-agent system for query processing with safety verification and critique built with Google A2A protocol, Google ADK, Llama Prompt Guard 2, Gemma 3 and Gemini 2.0 Flash.

Architecture

graph TD
    User[User/Client] <-->|Query/Response| Manager[Manager Agent]
    
    subgraph "A2A Protocol Communication"
        Manager <-->|Step 1: Check Query Safety| Safeguard[Safeguard Agent]
        Manager <-->|Step 2: Process Query| Processor[Processor Agent]
        Manager <-->|Step 3: Evaluate Response| Critic[Critic Agent]
    end
    
    Safeguard -->|Uses| Guard[Llama Prompt Guard 2]
    Processor -->|Uses| Gemma[Gemma 3 Model]
    Critic -->|Uses| Gemini[Gemini 2.0 Flash]
    
    classDef models fill:#f9f,stroke:#333,stroke-width:2px
    classDef agents fill:#bbf,stroke:#333,stroke-width:2px
    classDef external fill:#fff,stroke:#333,stroke-width:2px
    
    class Guard,Gemma,Gemini models
    class Manager,Safeguard,Processor,Critic agents
    class User external
    
    %% Flow description
    linkStyle 0 stroke:#000,stroke-width:2px
    linkStyle 1 stroke:#f00,stroke-width:2px
    linkStyle 2 stroke:#0a0,stroke-width:2px
    linkStyle 3 stroke:#00f,stroke-width:2px

This system uses four agents that work together:

1. Manager Agent: Coordinates the flow between agents
2. Safeguard Agent: Checks user queries for vulnerabilities using Meta's Prompt Guard 2 model
3. Processor Agent: Processes user queries using Gemma 3
4. Critic Agent: Evaluates responses for completeness and validity using Gemini 2.0 Flash

Each agent communicates using the Google Agent-to-Agent (A2A) Protocol for inter-agent communication, with FastAPI handling only external user-facing APIs.

Project Structure

The project is organized by agent, with each agent in its own module:

• agent_manager/: Manager Agent implementation
• agent_processor/: Processor Agent implementation
• agent_critic/: Critic Agent implemen