Stop using MCP for everything! (RAG vs. A2A vs. Sub-agents)

Key Concepts

• RAG (Retrieval Augmented Generation): Utilizing a knowledge base to enhance LLM responses with relevant context.
• MCP (Model Context Protocol): A standardized method for accessing data from various sources in a secure and controlled manner.
• A2A (Agent-to-Agent Protocol): Enabling communication and collaboration between AI agents from different organizations.
• Sub-Agents: Breaking down complex tasks into smaller, manageable components handled by specialized agents within a single application.
• ADK (Agent Development Kit): Google’s toolkit for building AI agents.
• Vortex AI Search/Rag Engine: Google Cloud product for indexing and retrieving information for RAG applications.
• Cloud Run/Agent Engine: Google Cloud platforms for deploying and running agent applications.
• Pub/Sub: Google Cloud’s messaging service used for asynchronous communication in A2A scenarios.

Choosing the Right AI Tool: A Deep Dive into MCP, A2A, RAG, and Sub-Agents

This discussion, featuring Debanchu from Google, clarifies the application of four key AI technologies – MCP, A2A, RAG, and sub-agents – using the example of a fictitious company, Fix It Fast, a network of home appliance repair technicians. The core argument is that while these tools appear complex, a logical framework exists for selecting the appropriate one based on the specific task.

1. RAG: The Technician Helper Bot – Accessing Knowledge

The first project, a technician helper bot designed to interpret error codes from older appliances, is a prime use case for RAG (Retrieval Augmented Generation). The need is to quickly present relevant information – the meaning of an error code – from a large library of repair manuals. Debanchu explains that RAG excels at providing specific snippets of information without requiring the technician to sift through entire documents.

Technical Implementation (Google Cloud): A mobile or web front-end would interface with a back-end application built using Google’s Agent Development Kit (ADK), running on Cloud Run or Agent Engine. The repair manuals would be indexed using Vortex AI Search (or a RAG engine), enabling efficient retrieval of relevant content.

2. MCP: The Executive Insights Dashboard – Federated Data Lookup

The second project, an executive dashboard requiring cross-departmental data analysis (revenue vs. marketing spend), demonstrates the utility of MCP (Model Context Protocol). MCP functions as a federated lookup tool, allowing an AI agent to query multiple databases without granting direct access.

Key Point: Debanchu emphasizes that using a single agent with access to all databases is impractical due to security and maintenance concerns. MCP allows individual teams (e.g., e-commerce, marketing) to maintain ownership and control over their data, exposing it through standardized MCP servers. Vortex AI acts as the reasoning engine, determining which MCP server to query.

Why not RAG? RAG requires access to the entire database, which may be a security risk or impractical. MCP provides controlled access to specific data points.

3. MCP (Action-Oriented): The Supply Bot – Taking Action in the Real World

The third project, a supply bot for ordering parts, re-introduces MCP, but this time for action rather than data retrieval. The agent uses MCP to trigger a backend system to place an order.

Distinction from API Calls: MCP provides a “standardized safe way for agents to discover and trigger business logic,” offering a more secure and manageable approach than traditional API calls. This allows the AI to interact with the real world in a controlled manner.

Technical Implementation: Similar to the technician helper bot, this utilizes a front-end, ADK-based back-end on Cloud Run, and an MCP server also running on Cloud Run.

4. Sub-Agents: The Triage and Dispatch System – Internal Teamwork

The fourth project, a new triage and dispatch system for incoming repair requests, highlights the benefits of sub-agents. The task – categorizing messages, prioritizing, scheduling, and responding – is too complex for a single prompt.

Methodology: The solution involves a “manager agent” delegating tasks to specialized sub-agents: one for diagnosis, one for scheduling, and one for technician notification. These sub-agents collaborate and share state using the ADK.

A2A vs. Sub-Agents: Debanchu clarifies that A2A (Agent-to-Agent Protocol) is designed for communication between different organizations, while sub-agents are more suitable for teamwork within a single application.

5. A2A: The Warranty Verifier – External Collaboration & Privacy

The final project, a warranty verifier, demonstrates the power of A2A (Agent-to-Agent Protocol). Fix It Fast lacks direct access to manufacturer databases due to privacy regulations.

Key Argument: A2A allows Fix It Fast’s agent to query a manufacturer’s agent, which acts as an “intelligent firewall,” handling private logic internally. This ensures data privacy and compliance. Communication can occur asynchronously via Pub/Sub. Human approval can be integrated for complex claims.

Why not MCP or Sub-Agents? MCP requires data access, which is restricted by privacy laws. Sub-agents operate within a single application and cannot directly access external databases.

Data & Statistics

While no specific numerical data was presented, the discussion implicitly highlights the increasing complexity of AI applications and the need for specialized tools to manage this complexity. The examples demonstrate the scalability benefits of using federated data access (MCP) and distributed task management (sub-agents).

Synthesis & Conclusion

Debanchu’s explanation provides a clear and actionable framework for choosing the right AI tool. The core takeaway is that the selection should be driven by the specific task:

• RAG: For knowledge retrieval.
• MCP: For standardized data access (lookup or action).
• Sub-Agents: For complex tasks requiring internal teamwork.
• A2A: For secure collaboration between different organizations.

The discussion emphasizes the importance of leveraging Google Cloud’s AI tooling (ADK, Cloud Run, Vortex AI, Pub/Sub) to build and deploy these solutions effectively. The concluding message, “Happy building,” underscores the potential for innovation in the rapidly evolving field of AI.