The Architect's Dilemma (www.oreilly.com)

AI architects are facing a practical design choice: expose capabilities as predictable, schema-driven tools (Model Context Protocol, MCP) or as stateful, goal-oriented agents (Agent2Agent, A2A). The essay reframes the decision as a UX-first question—“Who is calling your service?”—and introduces the Vending Machine vs Concierge metaphor. MCP is a vending machine: callers must provide a complete, machine-readable payload (e.g., a JSON-RPC tool call) and receive a deterministic result, which favors testability, auditability, and stateless transactions. A2A is a concierge: callers give high-level goals (“book a refundable flight under $450”), and the agent conducts a multi-turn dialog, captures preferences, maintains session state, and orchestrates processes over time. To validate the choice, the author offers a four-factor litmus test—determinism vs ambiguity, one-shot vs long-running processes, stateless vs stateful interactions, and direct control vs delegated ownership. These dimensions guide whether a service should present an MCP tool or an A2A agent; a complex internal workflow can still be an MCP if its external contract is a single deterministic call, while a simple but multi-turn user flow demands A2A. The piece also discusses common failure modes, testing implications, and patterns (like a Gatekeeper) for combining both interfaces so systems are both capable and reliable in real-world deployments.