1.3. The Interoperability Layer
💡 First Principle: Integrations grow as N×M unless a protocol makes them N+M. Without standards, every agent needs custom plumbing to every tool and every other agent — a combinatorial explosion. MCP and A2A are the two standards that collapse the explosion, and they split the problem cleanly: MCP connects an agent to resources; A2A connects an agent to peers.
You will meet both protocols repeatedly in Domain 2, the exam's heaviest — MCP as a tool type you attach to agents, A2A as a multi-agent integration path. Learning them here as protocols with distinct jobs pays off later, because exam distractors deliberately swap them: an MCP answer planted on an agent-collaboration question, an A2A answer planted on a tool-access question.
The comparison worth memorizing:
| Dimension | MCP (Model Context Protocol) | A2A (Agent-to-Agent) |
|---|---|---|
| Connects | One agent ↔ tools, data, resources | Peer agents ↔ each other |
| Relationship | Client consumes a server's capabilities | Agents negotiate and delegate tasks |
| Discovery | Server advertises tools dynamically in-session | Agent card describes capabilities |
| Mental model | USB-C port: plug capabilities in | Phone call between colleagues |
| In Copilot Studio | Add an MCP server as tools | Integrate external agents cross-platform |
| Typical exam cue | "expose our API's operations to the agent, updating automatically" | "our agent must work with an agent built on another vendor's platform" |
⚠️ Common Misconception: MCP and A2A solve the same problem. They are complementary layers: an agent might use MCP to reach a database and A2A to delegate a subtask to a partner company's agent — in the same conversation.
