Multi-agent Architecture

AI systems are moving from single-agent chat experiences to Level 4 architectures, where agents coordinate with other agents and external tools to complete real workflows. This workshop provides a practical blueprint for building multi-agent ecosystems using Agent-to-Agent (A2A) communication, the Model Context Protocol (MCP) for standardized tool connectivity, and Google’s Agent Development Kit (ADK) as a system design framework.

 

You’ll start with a fast, self-contained grounding in ReAct agents and the memory module (short-term, long-term, semantic memory and how it is commonly implemented with vector databases and session management). Then you’ll move into Level 4: how agents securely delegate tasks to other agents using A2A. 

 

The workshop then focuses on MCP as the standardized layer that helps agents connect to tools and data sources through a consistent protocol, including MCP’s core components (hosts, clients, servers, local data sources, remote services), how MCP differs from traditional APIs, and the limitations architects must design around. 

​

Finally, you’ll see how these pieces fit together through a multi-agent architecture to illustrate how to structure a production-style workflow that goes beyond a single reasoning loop.

​

Key Architectural Insights

• Levels of Agentic Architectures (1→4): How Level 4 shifts from single-agent reasoning to agent ecosystems.

• A2A Protocol in Practice: Secure agent collaboration using HTTP + JSON-RPC 2.0, supporting sequential/parallel/loop workflows and avoiding oversharing internal reasoning.

• MCP for Tool Connectivity: MCP components, what MCP standardizes, and why it changes agent-tool integration patterns.

• API vs MCP + Limitations: What MCP simplifies vs where it still fails

• ADK Mental Model: Agents as self-contained execution units; BaseAgent as the foundation and how ADK connects orchestration concepts.
   

 

Learning Outcomes

​

By the end of this workshop, attendees will be able to:

• Explain Level 4 agent ecosystems and when they outperform single-agent designs.

• Describe how A2A enables agent delegation and coordination (formats, workflow modes, security posture).

• Implement MCP as a structured connectivity layer and understand how it differs from conventional API tooling.

• Design around MCP’s practical limitations when building real systems.

• Map these pieces into an end-to-end multi-agent architecture