A practical breakdown of when to use MCP versus Google's A2A protocol for AI agent systems. MCP is designed for agent-to-tool communication (databases, APIs, files), while A2A handles agent-to-agent coordination where both sides have independent goals, state, and decision-making. The post walks through a real multi-agent customer support architecture that was incorrectly built on MCP alone, explains the four failure modes of wrapping agents in MCP servers (identity, long-running tasks, asymmetric authority, discovery), and describes A2A's three core primitives: Agent Cards, Tasks, and Artifacts. The recommended pattern is to use MCP as the internal toolbelt within each agent and A2A as the front door between agents. Practical guidance covers when each protocol is appropriate and how they compose in production systems.
