Building with AI · Building with the protocol stack
You can explain why MCP is not enough once you cross an agent boundary, what the A2A spec adds (capability discovery, task hand-off, return values), and when to reach for it instead of plain tool calls.
MCP handles the agent-to-tool relationship. A2A — Google's Agent2Agent protocol — handles the agent-to-agent relationship. Different problem, different shape.
A tool call is synchronous: the model emits arguments, a server returns a result, the model continues. An agent call is something else. The peer has its own loop, its own tools, its own memory, and possibly its own model. You don't tell it which steps to take. You hand it a goal and accept that the trajectory between hand-off and result belongs to it. A2A grew from 50 launch partners in April 2025 to 150+ partners in 2026, and AWS, Microsoft and Google all GA'd multi-agent orchestration on top of it.
This chapter does four things. First, when to even reach for a peer agent rather than a tool — the cost is real (15× single-agent token cost in Anthropic's own research benchmark). Then the A2A spec itself — Agent Cards, task hand-off, return values, capability discovery. Then a side-by-side with MCP so the two protocols stop blurring. Finally a worked example: an orchestrator dispatching three specialist agents and merging their work.
By the end you should know exactly when adding A2A pays for its complexity, and when a well-instrumented single loop is the right call.
Type: multi-choice
Chapter contains 4 lessons.