Building with AI · Building with the protocol stack
You can sketch a reference architecture that uses MCP for tools, A2A for delegation, and Skills for packaged capabilities — and explain which protocol each component speaks to which.
You have seen each layer in isolation. The last chapter wires them together. You will work through a reference architecture — one orchestrator, three specialist sub-agents, MCP servers for tools (Postgres, Sentry, a vector store), A2A for the orchestrator-to-specialist hand-off, Skills for the specialists' procedural knowledge — and then walk a real worked example: a customer-support workflow that ingests a ticket, routes, drafts, and escalates.
The point of the chapter is not to convince you to ship a five-agent system tomorrow. Most teams should not. Anthropic's own research warns that orchestration multiplies cost and surface area before it multiplies output. The point is to recognise the shape when you see it in production — and to know where the protocol seams are, so you can swap a specialist, replace a tool, or upgrade a Skill without rebuilding the system.
The closing lesson points at what to read and try next. Canonical examples, the MCP roadmap, the A2A RFC process, the Skills marketplace as it shapes up through 2027.
Type: multi-choice
Prompt: > In a reference architecture with an orchestrator and three specialist agents, which protocol does the orchestrator speak to a specialist?
Chapter contains 3 lessons.