(AWS re:invent 2025) Architecture Patterns for Strands and MCP

In the "Agentic Era," software architecture is defined by three components: 1. **Strands:** The Brain (Orchestration). 2. **MCP:** The Hands (Vertical Integration). 3. **A2A:** The Voice (Horizontal Collaboration). This guide explains when to use which protocol and how to implement them. ## The Mental Model: MCP vs. A2A The most common question is: "Why do we need two protocols?" | Feature | MCP (Model Context Protocol) | A2A (Agent-to-Agent) | | :--- | :--- | :--- | | **Direction** | Vertical (Agent $\to$ Data) | Horizontal (Agent $\leftrightarrow$ Agent) | | **Relationship** | Master/Slave | Peer-to-Peer | | **Transport** | JSON-RPC (Local/SSE) | JSON-RPC 2.0 (HTTP/Cloud 9000) | | **State** | Stateless | Stateful (Tasks) | | **Best For** | Database access, API calls | Delegation, Negotiation | **The Rule of Thumb:** Use MCP for *deterministic* tools (reading a file). Use A2A for *probabilistic* work (asking another agent to "review this plan"). ## Implementation: The Matryoshka Pattern A production agent is often a "Strands Agent" wrapped in an "A2A Server," consuming "MCP Clients." ```python # 1. Define the Core Agent (Strands) from strands import Agent from strands_tools.calculator import calculator strands_agent = Agent( model="anthropic.claude-3-5-sonnet-v2:0", tools=[calculator] # MCP Tools ) # 2. Wrap in A2A Server (Horizontal) from strands.multiagent.a2a import A2AServer a2a_server = A2AServer( agent=strands_agent, name="Finance Agent", port=9000 ) # 3. Serve a2a_server.serve() ``` ## Discovery: The Agent Card A2A agents don't use hardcoded IP addresses. They use **Semantic Discovery**. Every agent publishes an `agent-card.json`. ```json { "name": "TravelAgent", "description": "I can book flights and hotels.", "capabilities": ["book_flight", "check_visa"], "auth": { "type": "oauth2", "scopes": ["travel:write"] } } ``` When a user asks "I need a flight," the Orchestrator scans the network for agents with "travel" capabilities, reads their cards, and routes the task dynamically. ## Security: Zero Trust Mesh In a multi-agent system, we use **Identity Propagation**. When Agent A calls Agent B, it passes the original user's specific context (e.g., `User: Alice`). This ensures that even if Agent B has admin access to a database, it enforces permissions *as if* Alice were querying it directly. ## Conclusion Stop building monolithic agents. Build a **Composable Cognitive Enterprise** by using MCP for your data layer and A2A for your collaboration layer. **References:** * [Model Context Protocol](https://modelcontextprotocol.io)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

# 1. Define the Core Agent (Strands)
from strands import Agent
from strands_tools.calculator import calculator

strands_agent = Agent(
    model="anthropic.claude-3-5-sonnet-v2:0",
    tools=[calculator] # MCP Tools
)

# 2. Wrap in A2A Server (Horizontal)
from strands.multiagent.a2a import A2AServer

a2a_server = A2AServer(
    agent=strands_agent,
    name="Finance Agent",
    port=9000
)

# 3. Serve
a2a_server.serve()

1
2
3
4
5
6
7
8
9

{
  "name": "TravelAgent",
  "description": "I can book flights and hotels.",
  "capabilities": ["book_flight", "check_visa"],
  "auth": {
    "type": "oauth2",
    "scopes": ["travel:write"]
  }
}