Glonce

MCP, A2A and ACP handshakes benchmarked: round trips dominate cost

researchThe Glonce editorsarXiv🔥58Glonce rating58Impact45Novelty60Relevance90Surprise35Players25Credibility70

Every exchange between AI agents starts with a handshake: the initiating agent has to find its counterpart, learn what it can do, and set up a session before the first useful message goes out. The authors of a new arXiv paper say surveys of agent communication protocols already name handshake overhead as an efficiency criterion, but none report measurements. Their paper claims to be the first empirical, phase-decomposed benchmark of that handshake.

The study covers three major agent protocols: MCP, in both its 2025 "initialize" era and its 2026 "discover" era, plus A2A and ACP. Using each protocol's official SDK, the authors instrument every request for round trips, per-phase latency, wire bytes and token cost. They run four scenarios: cold-start, warm-repeat, capability-mismatch and degraded-metadata. Each cell has N=50 runs, measured on loopback and under an emulated 50 ms round-trip time (RTT).

The headline finding is that round-trip count dominates handshake cost. A legacy MCP handshake takes three round trips and 230 ms at 50 ms RTT. A2A needs one round trip and 61.0 ms.

Round trips are not the whole bill, though. When an LLM reads the counterpart's metadata, handshake cost scales with the counterpart's capability inventory, at 24 to 107 tokens per capability. At that rate, a counterpart that advertises 200 capabilities uses 11% of a 200k-token context window before the agent reads its own task.

The authors also point to MCP's own history as support. The 2026 revision of MCP cuts its handshake from three round trips to two and, with era pinning, to zero on warm reconnects. They describe this protocol evolution as corroborating the cost they measure. The full harness, raw data and analysis are released on GitHub at github.com/pvarshh/handshake-benchmark.

Key facts

  • The paper claims the first empirical, phase-decomposed benchmark of the handshake in MCP (2025 initialize and 2026 discover eras), A2A and ACP.
  • A legacy MCP handshake takes three round trips and 230 ms at 50 ms emulated RTT; A2A needs one round trip and 61.0 ms.
  • When an LLM reads a counterpart's metadata, cost is 24 to 107 tokens per capability, so a 200-capability counterpart takes 11% of a 200k-token context window.
  • MCP's 2026 revision cuts its handshake from three round trips to two, and to zero on warm reconnects with era pinning.
  • Each cell uses N=50 runs on loopback and under 50 ms RTT; the harness, raw data and analysis are on GitHub.

Why it matters

Agent protocols are usually compared on features. This paper looks at a cost every agent pays before doing any work: the handshake. The authors say earlier surveys list handshake overhead as an efficiency criterion without reporting any measurements, so this is offered as the first set of numbers. The result is concrete: round-trip count drives the cost, and MCP's own move from three round trips to two, and to zero on warm reconnects with era pinning, points the same way.

Who it affects

People who build or pick between agent protocols, namely MCP, A2A and ACP, are the direct audience. So are teams running agents that connect to many counterparts, since the token cost grows with the size of the counterpart's capability inventory. Anyone whose agents reconnect often is affected by the cold versus warm distinction the study measures.

How to use it

The authors release the full harness, raw data and analysis at github.com/pvarshh/handshake-benchmark, so the measurements can be inspected or rerun. The two cost levers the paper identifies are fewer round trips and smaller capability inventories: at 24 to 107 tokens per capability, the amount of metadata an LLM reads at handshake time adds directly to context use.

How solid is it

The setup is clearly described: official SDKs for each protocol, four scenarios, N=50 per cell, and two network conditions (loopback and an emulated 50 ms RTT). Data and code are public. The source is an arXiv abstract, so these are the authors' own claims, including the claim to be first. The abstract does not say which condition the A2A 61.0 ms figure refers to, though it follows the 50 ms RTT figure for legacy MCP.

Risks and caveats

The abstract gives millisecond figures only for legacy MCP and A2A; none for ACP or for the 2026 MCP discover era. The 24 to 107 tokens per capability range is not broken down by protocol. The 11% figure is for a hypothetical 200-capability counterpart and a 200k-token window, not a named model or real deployment. The abstract does not say which protocol is best overall or make a recommendation.

“Round-trip count dominates handshake cost.”

— From the paper's abstract