The Cost of Saying Hello: Measuring Handshake Overhead in AI Agent Communication Protocols

Every interaction between AI agents begins with a handshake: the initiator must locate its counterpart, learn what it can do, and establish a session before the first useful message. Surveys of agent communication protocols name handshake overhead as an efficiency criterion, yet none report measurements. We decompose the handshake into four phases and measure it in three protocols using their official SDKs: MCP (in both its 2025 initialize and 2026 discover eras), A2A, and ACP. Round-trip count dominates. A legacy MCP handshake costs 230 ms at 50 ms RTT against 61.0 ms for A2A, and the legacy client opens a fresh TCP connection per JSON-RPC exchange, a behaviour we traced to the SDK and reported upstream. Round trips are not the whole bill: when an LLM reads the counterpart's metadata, cost scales with inventory size at 24-107 tokens per capability, so a 200-capability counterpart spends 11% of a 200k-token context window before the agent reads its own task. MCP's own 2026 revision cuts three round trips to two, and to zero on pinned warm reconnects, protocol evolution that corroborates the cost we measure. We release the harness, raw per-run data, and analysis.