The Pulse
New Benchmark Puts a Latency Price on Agent Handshakes
Pranav Varshney’s arXiv benchmark compares handshake overhead in MCP, A2A and ACP agent protocols. It finds that round trips drive cold-start latency, while caching can sharply reduce the cost of repeat connections.
AI.info Team · Written with AI · fact-checked ·
A legacy Model Context Protocol (MCP) handshake takes 230 milliseconds in a new benchmark; an Agent2Agent (A2A) connection takes 61 milliseconds. But the comparison has a wrinkle: a newer MCP design cuts the measured cold-start time to 122 milliseconds, and caching can remove network exchanges entirely for repeat connections. The results, published by Pranav Varshney on arXiv on August 12, put numbers on a cost that agent-protocol comparisons have often treated as a design criterion without measuring.
“Round-trip count dominates handshake cost.”
Pranav Varshney, “The Cost of Saying Hello: An Empirical Study of Handshake Overhead in AI Agent Communication Protocols”
Three round trips versus one
A handshake is the exchange that lets an agent find another system, learn what it can do and establish that it is ready to receive a task. Varshney measured that setup in MCP, A2A and the Agent Communication Protocol (ACP), using each protocol’s official Python SDK and a minimal counterpart exposing one capability. The benchmark tracks application-level round trips, latency, bytes and metadata tokens across cold starts, repeat connections, capability mismatches and degraded conditions.
For cold starts, the study reports median latency of 230 milliseconds for legacy MCP, 122 milliseconds for the 2026 MCP revision, 61.0 milliseconds for A2A and 62.1 milliseconds for ACP under an emulated 50-millisecond round-trip time. The tested handshakes took three application round trips for legacy MCP, two for modern MCP, and one each for A2A and ACP. Those measurements are SDK-and-configuration results, not a universal ranking of every implementation.
MCP’s newer handshake narrows the gap
The contrast between MCP’s two versions complicates any simple verdict. The paper says the newer revision replaces the older three-message initialization with a discovery probe and cacheable results; in the benchmark, that change cuts cold-start latency by 47 percent compared with legacy MCP. With the SDK’s default repeat-connection policy, modern MCP still makes one round trip and takes a median 63.5 milliseconds. When the client pins the counterpart’s protocol era and reuses cached metadata, it reaches readiness without a network exchange; A2A and ACP also reach zero-round-trip warm paths when their cards or manifests are cached.
Those gains depend on the connection being familiar. A cold connection to a stranger still has to fetch and inspect capability information, while cached descriptions can become stale. The paper also found that a capability mismatch costs the full handshake before the tested clients can reject a counterpart: at 50 milliseconds of emulated RTT, that took 231 milliseconds for legacy MCP and about 58 to 62 milliseconds for A2A and ACP.
Capability lists carry a separate token bill
Network latency is only one part of the setup cost when an LLM reads the other agent’s metadata. In a test that expanded capability inventories to 200 entries, the paper counts 4,891 tokens for A2A, 15,618 for MCP and 21,405 for ACP. The ACP figure is about 11 percent of a 200,000-token context window; the same percentage does not apply across protocols. Varshney estimates marginal metadata costs from 24 tokens per capability for A2A to 107 for ACP, with MCP at 78.
The token estimates come from replicated entries in captured metadata, counted with the o200k_base tokenizer; they are not observations of 200-capability agents in production. The author also notes that the experiment uses anonymous configurations and excludes authentication, which could add further exchanges. Its 50-millisecond network condition is emulated at the application layer, and the reported figures exclude real-network TCP and TLS setup.
A benchmark of protocol setup, not task performance
ACP remains in the comparison even though its project merged into A2A in 2025, because deployments persist and its one-round-trip design gives the study another point of comparison. The paper measures only the connection phase, not how well agents complete tasks once connected. Varshney says the benchmark’s harness, raw per-run data and analysis are available alongside the paper, allowing teams to inspect or reproduce measurements using the tested SDK versions.
The clearest result is narrower than a protocol-wide winner: on the tested cold path, fewer exchanges meant less waiting, while reusable metadata made repeat connections far cheaper. For a 200-entry capability list, the measured metadata ranged from 4,891 tokens in A2A to 21,405 in ACP.